xref: /freebsd/sys/dev/mlx4/mlx4_en/mlx4_en_rx.c (revision ce6a89e27cd190313be39bb479880aeda4778436)
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 #if (MLX4_EN_MAX_RX_SEGS == 1)
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 =
53 	    ((struct mlx4_en_rx_desc *)ring->buf) + index;
54 	int i;
55 
56 	/* Set size and memtype fields */
57 	rx_desc->data[0].byte_count = cpu_to_be32(priv->rx_mb_size - MLX4_NET_IP_ALIGN);
58 	rx_desc->data[0].lkey = cpu_to_be32(priv->mdev->mr.key);
59 
60 	/*
61 	 * If the number of used fragments does not fill up the ring
62 	 * stride, remaining (unused) fragments must be padded with
63 	 * null address/size and a special memory key:
64 	 */
65 	for (i = 1; i < MLX4_EN_MAX_RX_SEGS; i++) {
66 		rx_desc->data[i].byte_count = 0;
67 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
68 		rx_desc->data[i].addr = 0;
69 	}
70 }
71 #endif
72 
73 static inline struct mbuf *
74 mlx4_en_alloc_mbuf(struct mlx4_en_rx_ring *ring)
75 {
76 	struct mbuf *mb;
77 
78 #if (MLX4_EN_MAX_RX_SEGS == 1)
79         mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, ring->rx_mb_size);
80         if (likely(mb != NULL))
81 		mb->m_pkthdr.len = mb->m_len = ring->rx_mb_size;
82 #else
83 	mb = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, MLX4_EN_MAX_RX_BYTES);
84 	if (likely(mb != NULL)) {
85 		struct mbuf *mb_head = mb;
86 		int i;
87 
88 		mb->m_len = MLX4_EN_MAX_RX_BYTES;
89 		mb->m_pkthdr.len = MLX4_EN_MAX_RX_BYTES;
90 
91 		for (i = 1; i != MLX4_EN_MAX_RX_SEGS; i++) {
92 			if (mb_head->m_pkthdr.len >= ring->rx_mb_size)
93 				break;
94 			mb = (mb->m_next = m_getjcl(M_NOWAIT, MT_DATA, 0, MLX4_EN_MAX_RX_BYTES));
95 			if (unlikely(mb == NULL)) {
96 				m_freem(mb_head);
97 				return (NULL);
98 			}
99 			mb->m_len = MLX4_EN_MAX_RX_BYTES;
100 			mb_head->m_pkthdr.len += MLX4_EN_MAX_RX_BYTES;
101 		}
102 		/* rewind to first mbuf in chain */
103 		mb = mb_head;
104 	}
105 #endif
106 	return (mb);
107 }
108 
109 static int
110 mlx4_en_alloc_buf(struct mlx4_en_rx_ring *ring, struct mlx4_en_rx_desc *rx_desc,
111     struct mlx4_en_rx_mbuf *mb_list)
112 {
113 	bus_dma_segment_t segs[MLX4_EN_MAX_RX_SEGS];
114 	bus_dmamap_t map;
115 	struct mbuf *mb;
116 	int nsegs;
117 	int err;
118 #if (MLX4_EN_MAX_RX_SEGS != 1)
119 	int i;
120 #endif
121 
122 	/* try to allocate a new spare mbuf */
123 	if (unlikely(ring->spare.mbuf == NULL)) {
124 		mb = mlx4_en_alloc_mbuf(ring);
125 		if (unlikely(mb == NULL))
126 			return (-ENOMEM);
127 
128 		/* make sure IP header gets aligned */
129 		m_adj(mb, MLX4_NET_IP_ALIGN);
130 
131 		/* load spare mbuf into BUSDMA */
132 		err = -bus_dmamap_load_mbuf_sg(ring->dma_tag, ring->spare.dma_map,
133 		    mb, ring->spare.segs, &nsegs, BUS_DMA_NOWAIT);
134 		if (unlikely(err != 0)) {
135 			m_freem(mb);
136 			return (err);
137 		}
138 
139 		/* store spare info */
140 		ring->spare.mbuf = mb;
141 
142 #if (MLX4_EN_MAX_RX_SEGS != 1)
143 		/* zero remaining segs */
144 		for (i = nsegs; i != MLX4_EN_MAX_RX_SEGS; i++) {
145 			ring->spare.segs[i].ds_addr = 0;
146 			ring->spare.segs[i].ds_len = 0;
147 		}
148 #endif
149 		bus_dmamap_sync(ring->dma_tag, ring->spare.dma_map,
150 		    BUS_DMASYNC_PREREAD);
151 	}
152 
153 	/* synchronize and unload the current mbuf, if any */
154 	if (likely(mb_list->mbuf != NULL)) {
155 		bus_dmamap_sync(ring->dma_tag, mb_list->dma_map,
156 		    BUS_DMASYNC_POSTREAD);
157 		bus_dmamap_unload(ring->dma_tag, mb_list->dma_map);
158 	}
159 
160 	mb = mlx4_en_alloc_mbuf(ring);
161 	if (unlikely(mb == NULL))
162 		goto use_spare;
163 
164 	/* make sure IP header gets aligned */
165 	m_adj(mb, MLX4_NET_IP_ALIGN);
166 
167 	err = -bus_dmamap_load_mbuf_sg(ring->dma_tag, mb_list->dma_map,
168 	    mb, segs, &nsegs, BUS_DMA_NOWAIT);
169 	if (unlikely(err != 0)) {
170 		m_freem(mb);
171 		goto use_spare;
172 	}
173 
174 #if (MLX4_EN_MAX_RX_SEGS == 1)
175 	rx_desc->data[0].addr = cpu_to_be64(segs[0].ds_addr);
176 #else
177 	for (i = 0; i != nsegs; i++) {
178 		rx_desc->data[i].byte_count = cpu_to_be32(segs[i].ds_len);
179 		rx_desc->data[i].lkey = ring->rx_mr_key_be;
180 		rx_desc->data[i].addr = cpu_to_be64(segs[i].ds_addr);
181 	}
182 	for (; i != MLX4_EN_MAX_RX_SEGS; i++) {
183 		rx_desc->data[i].byte_count = 0;
184 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
185 		rx_desc->data[i].addr = 0;
186 	}
187 #endif
188 	mb_list->mbuf = mb;
189 
190 	bus_dmamap_sync(ring->dma_tag, mb_list->dma_map, BUS_DMASYNC_PREREAD);
191 	return (0);
192 
193 use_spare:
194 	/* swap DMA maps */
195 	map = mb_list->dma_map;
196 	mb_list->dma_map = ring->spare.dma_map;
197 	ring->spare.dma_map = map;
198 
199 	/* swap MBUFs */
200 	mb_list->mbuf = ring->spare.mbuf;
201 	ring->spare.mbuf = NULL;
202 
203 	/* store physical address */
204 #if (MLX4_EN_MAX_RX_SEGS == 1)
205 	rx_desc->data[0].addr = cpu_to_be64(ring->spare.segs[0].ds_addr);
206 #else
207 	for (i = 0; i != MLX4_EN_MAX_RX_SEGS; i++) {
208 		if (ring->spare.segs[i].ds_len != 0) {
209 			rx_desc->data[i].byte_count = cpu_to_be32(ring->spare.segs[i].ds_len);
210 			rx_desc->data[i].lkey = ring->rx_mr_key_be;
211 			rx_desc->data[i].addr = cpu_to_be64(ring->spare.segs[i].ds_addr);
212 		} else {
213 			rx_desc->data[i].byte_count = 0;
214 			rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
215 			rx_desc->data[i].addr = 0;
216 		}
217 	}
218 #endif
219 	return (0);
220 }
221 
222 static void
223 mlx4_en_free_buf(struct mlx4_en_rx_ring *ring, struct mlx4_en_rx_mbuf *mb_list)
224 {
225 	bus_dmamap_t map = mb_list->dma_map;
226 	bus_dmamap_sync(ring->dma_tag, map, BUS_DMASYNC_POSTREAD);
227 	bus_dmamap_unload(ring->dma_tag, map);
228 	m_freem(mb_list->mbuf);
229 	mb_list->mbuf = NULL;	/* safety clearing */
230 }
231 
232 static int
233 mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
234     struct mlx4_en_rx_ring *ring, int index)
235 {
236 	struct mlx4_en_rx_desc *rx_desc =
237 	    ((struct mlx4_en_rx_desc *)ring->buf) + index;
238 	struct mlx4_en_rx_mbuf *mb_list = ring->mbuf + index;
239 
240 	mb_list->mbuf = NULL;
241 
242 	if (mlx4_en_alloc_buf(ring, rx_desc, mb_list)) {
243 		priv->port_stats.rx_alloc_failed++;
244 		return (-ENOMEM);
245 	}
246 	return (0);
247 }
248 
249 static inline void
250 mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
251 {
252 	*ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
253 }
254 
255 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
256 {
257 	struct mlx4_en_rx_ring *ring;
258 	int ring_ind;
259 	int buf_ind;
260 	int new_size;
261 	int err;
262 
263 	for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
264 		for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
265 			ring = priv->rx_ring[ring_ind];
266 
267 			err = mlx4_en_prepare_rx_desc(priv, ring,
268 						      ring->actual_size);
269 			if (err) {
270 				if (ring->actual_size == 0) {
271 					en_err(priv, "Failed to allocate "
272 						     "enough rx buffers\n");
273 					return -ENOMEM;
274 				} else {
275 					new_size =
276 						rounddown_pow_of_two(ring->actual_size);
277 					en_warn(priv, "Only %d buffers allocated "
278 						      "reducing ring size to %d\n",
279 						ring->actual_size, new_size);
280 					goto reduce_rings;
281 				}
282 			}
283 			ring->actual_size++;
284 			ring->prod++;
285 		}
286 	}
287 	return 0;
288 
289 reduce_rings:
290 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
291 		ring = priv->rx_ring[ring_ind];
292 		while (ring->actual_size > new_size) {
293 			ring->actual_size--;
294 			ring->prod--;
295 			mlx4_en_free_buf(ring,
296 			    ring->mbuf + ring->actual_size);
297 		}
298 	}
299 
300 	return 0;
301 }
302 
303 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
304 				struct mlx4_en_rx_ring *ring)
305 {
306 	int index;
307 
308 	en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
309 	       ring->cons, ring->prod);
310 
311 	/* Unmap and free Rx buffers */
312 	BUG_ON((u32) (ring->prod - ring->cons) > ring->actual_size);
313 	while (ring->cons != ring->prod) {
314 		index = ring->cons & ring->size_mask;
315 		en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
316 		mlx4_en_free_buf(ring, ring->mbuf + index);
317 		++ring->cons;
318 	}
319 }
320 
321 void mlx4_en_set_num_rx_rings(struct mlx4_en_dev *mdev)
322 {
323 	int i;
324 	int num_of_eqs;
325 	int num_rx_rings;
326 	struct mlx4_dev *dev = mdev->dev;
327 
328 	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH) {
329 		num_of_eqs = max_t(int, MIN_RX_RINGS,
330 				   min_t(int,
331 					 mlx4_get_eqs_per_port(mdev->dev, i),
332 					 DEF_RX_RINGS));
333 
334 		num_rx_rings = mlx4_low_memory_profile() ? MIN_RX_RINGS :
335 							   num_of_eqs;
336 		mdev->profile.prof[i].rx_ring_num =
337 			rounddown_pow_of_two(num_rx_rings);
338 	}
339 }
340 
341 void mlx4_en_calc_rx_buf(struct net_device *dev)
342 {
343 	struct mlx4_en_priv *priv = netdev_priv(dev);
344 	int eff_mtu = dev->if_mtu + ETH_HLEN + VLAN_HLEN + ETH_FCS_LEN +
345 	    MLX4_NET_IP_ALIGN;
346 
347 	if (eff_mtu > MJUM16BYTES) {
348 		en_err(priv, "MTU(%u) is too big\n", (unsigned)dev->if_mtu);
349                 eff_mtu = MJUM16BYTES;
350         } else if (eff_mtu > MJUM9BYTES) {
351                 eff_mtu = MJUM16BYTES;
352         } else if (eff_mtu > MJUMPAGESIZE) {
353                 eff_mtu = MJUM9BYTES;
354         } else if (eff_mtu > MCLBYTES) {
355                 eff_mtu = MJUMPAGESIZE;
356         } else {
357                 eff_mtu = MCLBYTES;
358         }
359 
360 	priv->rx_mb_size = eff_mtu;
361 
362 	en_dbg(DRV, priv, "Effective RX MTU: %d bytes\n", eff_mtu);
363 }
364 
365 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
366 			   struct mlx4_en_rx_ring **pring,
367 			   u32 size, int node)
368 {
369 	struct mlx4_en_dev *mdev = priv->mdev;
370 	struct mlx4_en_rx_ring *ring;
371 	int err;
372 	int tmp;
373 	uint32_t x;
374 
375         ring = kzalloc(sizeof(struct mlx4_en_rx_ring), GFP_KERNEL);
376         if (!ring) {
377                 en_err(priv, "Failed to allocate RX ring structure\n");
378                 return -ENOMEM;
379         }
380 
381 	/* Create DMA descriptor TAG */
382 	if ((err = -bus_dma_tag_create(
383 	    bus_get_dma_tag(mdev->pdev->dev.bsddev),
384 	    1,				/* any alignment */
385 	    0,				/* no boundary */
386 	    BUS_SPACE_MAXADDR,		/* lowaddr */
387 	    BUS_SPACE_MAXADDR,		/* highaddr */
388 	    NULL, NULL,			/* filter, filterarg */
389 	    MJUM16BYTES,		/* maxsize */
390 	    MLX4_EN_MAX_RX_SEGS,	/* nsegments */
391 	    MJUM16BYTES,		/* maxsegsize */
392 	    0,				/* flags */
393 	    NULL, NULL,			/* lockfunc, lockfuncarg */
394 	    &ring->dma_tag))) {
395 		en_err(priv, "Failed to create DMA tag\n");
396 		goto err_ring;
397 	}
398 
399 	ring->prod = 0;
400 	ring->cons = 0;
401 	ring->size = size;
402 	ring->size_mask = size - 1;
403 
404 	ring->log_stride = ilog2(sizeof(struct mlx4_en_rx_desc));
405 	ring->buf_size = (ring->size * sizeof(struct mlx4_en_rx_desc)) + TXBB_SIZE;
406 
407 	tmp = size * sizeof(struct mlx4_en_rx_mbuf);
408 
409         ring->mbuf = kzalloc(tmp, GFP_KERNEL);
410         if (ring->mbuf == NULL) {
411                 err = -ENOMEM;
412                 goto err_dma_tag;
413         }
414 
415 	err = -bus_dmamap_create(ring->dma_tag, 0, &ring->spare.dma_map);
416 	if (err != 0)
417 		goto err_info;
418 
419 	for (x = 0; x != size; x++) {
420 		err = -bus_dmamap_create(ring->dma_tag, 0,
421 		    &ring->mbuf[x].dma_map);
422 		if (err != 0) {
423 			while (x--)
424 				bus_dmamap_destroy(ring->dma_tag,
425 				    ring->mbuf[x].dma_map);
426 			goto err_info;
427 		}
428 	}
429 	en_dbg(DRV, priv, "Allocated MBUF ring at addr:%p size:%d\n",
430 		 ring->mbuf, tmp);
431 
432 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
433 				 ring->buf_size, 2 * PAGE_SIZE);
434 	if (err)
435 		goto err_dma_map;
436 
437 	err = mlx4_en_map_buffer(&ring->wqres.buf);
438 	if (err) {
439 		en_err(priv, "Failed to map RX buffer\n");
440 		goto err_hwq;
441 	}
442 	ring->buf = ring->wqres.buf.direct.buf;
443 	*pring = ring;
444 	return 0;
445 
446 err_hwq:
447 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
448 err_dma_map:
449 	for (x = 0; x != size; x++) {
450 		bus_dmamap_destroy(ring->dma_tag,
451 		    ring->mbuf[x].dma_map);
452 	}
453 	bus_dmamap_destroy(ring->dma_tag, ring->spare.dma_map);
454 err_info:
455 	vfree(ring->mbuf);
456 err_dma_tag:
457 	bus_dma_tag_destroy(ring->dma_tag);
458 err_ring:
459 	kfree(ring);
460 	return (err);
461 }
462 
463 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
464 {
465 	struct mlx4_en_rx_ring *ring;
466 #if (MLX4_EN_MAX_RX_SEGS == 1)
467 	int i;
468 #endif
469 	int ring_ind;
470 	int err;
471 
472 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
473 		ring = priv->rx_ring[ring_ind];
474 
475 		ring->prod = 0;
476 		ring->cons = 0;
477 		ring->actual_size = 0;
478 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
479                 ring->rx_mb_size = priv->rx_mb_size;
480 
481 		if (sizeof(struct mlx4_en_rx_desc) <= TXBB_SIZE) {
482 			/* Stamp first unused send wqe */
483 			__be32 *ptr = (__be32 *)ring->buf;
484 			__be32 stamp = cpu_to_be32(1 << STAMP_SHIFT);
485 			*ptr = stamp;
486 			/* Move pointer to start of rx section */
487 			ring->buf += TXBB_SIZE;
488 		}
489 
490 		ring->log_stride = ilog2(sizeof(struct mlx4_en_rx_desc));
491 		ring->buf_size = ring->size * sizeof(struct mlx4_en_rx_desc);
492 
493 		memset(ring->buf, 0, ring->buf_size);
494 		mlx4_en_update_rx_prod_db(ring);
495 
496 #if (MLX4_EN_MAX_RX_SEGS == 1)
497 		/* Initialize all descriptors */
498 		for (i = 0; i < ring->size; i++)
499 			mlx4_en_init_rx_desc(priv, ring, i);
500 #endif
501 		ring->rx_mr_key_be = cpu_to_be32(priv->mdev->mr.key);
502 
503 #ifdef INET
504 		/* Configure lro mngr */
505 		if (priv->dev->if_capenable & IFCAP_LRO) {
506 			if (tcp_lro_init(&ring->lro))
507 				priv->dev->if_capenable &= ~IFCAP_LRO;
508 			else
509 				ring->lro.ifp = priv->dev;
510 		}
511 #endif
512 	}
513 
514 
515 	err = mlx4_en_fill_rx_buffers(priv);
516 	if (err)
517 		goto err_buffers;
518 
519 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
520 		ring = priv->rx_ring[ring_ind];
521 
522 		ring->size_mask = ring->actual_size - 1;
523 		mlx4_en_update_rx_prod_db(ring);
524 	}
525 
526 	return 0;
527 
528 err_buffers:
529 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
530 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
531 
532 	ring_ind = priv->rx_ring_num - 1;
533 
534 	while (ring_ind >= 0) {
535 		ring = priv->rx_ring[ring_ind];
536 		if (sizeof(struct mlx4_en_rx_desc) <= TXBB_SIZE)
537 			ring->buf -= TXBB_SIZE;
538 		ring_ind--;
539 	}
540 
541 	return err;
542 }
543 
544 
545 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
546 			     struct mlx4_en_rx_ring **pring,
547 			     u32 size)
548 {
549 	struct mlx4_en_dev *mdev = priv->mdev;
550 	struct mlx4_en_rx_ring *ring = *pring;
551 	uint32_t x;
552 
553 	mlx4_en_unmap_buffer(&ring->wqres.buf);
554 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * sizeof(struct mlx4_en_rx_desc) + TXBB_SIZE);
555 	for (x = 0; x != size; x++)
556 		bus_dmamap_destroy(ring->dma_tag, ring->mbuf[x].dma_map);
557 	/* free spare mbuf, if any */
558 	if (ring->spare.mbuf != NULL) {
559 		bus_dmamap_sync(ring->dma_tag, ring->spare.dma_map,
560 		    BUS_DMASYNC_POSTREAD);
561 		bus_dmamap_unload(ring->dma_tag, ring->spare.dma_map);
562 		m_freem(ring->spare.mbuf);
563 	}
564 	bus_dmamap_destroy(ring->dma_tag, ring->spare.dma_map);
565 	vfree(ring->mbuf);
566 	bus_dma_tag_destroy(ring->dma_tag);
567 	kfree(ring);
568 	*pring = NULL;
569 #ifdef CONFIG_RFS_ACCEL
570 	mlx4_en_cleanup_filters(priv, ring);
571 #endif
572 }
573 
574 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
575 				struct mlx4_en_rx_ring *ring)
576 {
577 #ifdef INET
578 	tcp_lro_free(&ring->lro);
579 #endif
580 	mlx4_en_free_rx_buf(priv, ring);
581 	if (sizeof(struct mlx4_en_rx_desc) <= TXBB_SIZE)
582 		ring->buf -= TXBB_SIZE;
583 }
584 
585 
586 static void validate_loopback(struct mlx4_en_priv *priv, struct mbuf *mb)
587 {
588 	int i;
589 	int offset = ETHER_HDR_LEN;
590 
591 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
592 		if (*(mb->m_data + offset) != (unsigned char) (i & 0xff))
593 			goto out_loopback;
594 	}
595 	/* Loopback found */
596 	priv->loopback_ok = 1;
597 
598 out_loopback:
599 	m_freem(mb);
600 }
601 
602 
603 static inline int invalid_cqe(struct mlx4_en_priv *priv,
604 			      struct mlx4_cqe *cqe)
605 {
606 	/* Drop packet on bad receive or bad checksum */
607 	if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
608 		     MLX4_CQE_OPCODE_ERROR)) {
609 		en_err(priv, "CQE completed in error - vendor syndrom:%d syndrom:%d\n",
610 		       ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
611 		       ((struct mlx4_err_cqe *)cqe)->syndrome);
612 		return 1;
613 	}
614 	if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
615 		en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
616 		return 1;
617 	}
618 
619 	return 0;
620 }
621 
622 static struct mbuf *
623 mlx4_en_rx_mb(struct mlx4_en_priv *priv, struct mlx4_en_rx_ring *ring,
624     struct mlx4_en_rx_desc *rx_desc, struct mlx4_en_rx_mbuf *mb_list,
625     int length)
626 {
627 #if (MLX4_EN_MAX_RX_SEGS != 1)
628 	struct mbuf *mb_head;
629 #endif
630 	struct mbuf *mb;
631 
632 	/* optimise reception of small packets */
633 	if (length <= (MHLEN - MLX4_NET_IP_ALIGN) &&
634 	    (mb = m_gethdr(M_NOWAIT, MT_DATA)) != NULL) {
635 
636 		/* set packet length */
637 		mb->m_pkthdr.len = mb->m_len = length;
638 
639 		/* make sure IP header gets aligned */
640 		mb->m_data += MLX4_NET_IP_ALIGN;
641 
642 		bus_dmamap_sync(ring->dma_tag, mb_list->dma_map,
643 		    BUS_DMASYNC_POSTREAD);
644 
645 		bcopy(mtod(mb_list->mbuf, caddr_t), mtod(mb, caddr_t), length);
646 
647 		return (mb);
648 	}
649 
650 	/* get mbuf */
651 	mb = mb_list->mbuf;
652 
653 	/* collect used fragment while atomically replacing it */
654 	if (mlx4_en_alloc_buf(ring, rx_desc, mb_list))
655 		return (NULL);
656 
657 	/* range check hardware computed value */
658 	if (unlikely(length > mb->m_pkthdr.len))
659 		length = mb->m_pkthdr.len;
660 
661 #if (MLX4_EN_MAX_RX_SEGS == 1)
662 	/* update total packet length in packet header */
663 	mb->m_len = mb->m_pkthdr.len = length;
664 #else
665 	mb->m_pkthdr.len = length;
666 	for (mb_head = mb; mb != NULL; mb = mb->m_next) {
667 		if (mb->m_len > length)
668 			mb->m_len = length;
669 		length -= mb->m_len;
670 		if (likely(length == 0)) {
671 			if (likely(mb->m_next != NULL)) {
672 				/* trim off empty mbufs */
673 				m_freem(mb->m_next);
674 				mb->m_next = NULL;
675 			}
676 			break;
677 		}
678 	}
679 	/* rewind to first mbuf in chain */
680 	mb = mb_head;
681 #endif
682 	return (mb);
683 }
684 
685 static __inline int
686 mlx4_en_rss_hash(__be16 status, int udp_rss)
687 {
688 	enum {
689 		status_all = cpu_to_be16(
690 			MLX4_CQE_STATUS_IPV4    |
691 			MLX4_CQE_STATUS_IPV4F   |
692 			MLX4_CQE_STATUS_IPV6    |
693 			MLX4_CQE_STATUS_TCP     |
694 			MLX4_CQE_STATUS_UDP),
695 		status_ipv4_tcp = cpu_to_be16(
696 			MLX4_CQE_STATUS_IPV4    |
697 			MLX4_CQE_STATUS_TCP),
698 		status_ipv6_tcp = cpu_to_be16(
699 			MLX4_CQE_STATUS_IPV6    |
700 			MLX4_CQE_STATUS_TCP),
701 		status_ipv4_udp = cpu_to_be16(
702 			MLX4_CQE_STATUS_IPV4    |
703 			MLX4_CQE_STATUS_UDP),
704 		status_ipv6_udp = cpu_to_be16(
705 			MLX4_CQE_STATUS_IPV6    |
706 			MLX4_CQE_STATUS_UDP),
707 		status_ipv4 = cpu_to_be16(MLX4_CQE_STATUS_IPV4),
708 		status_ipv6 = cpu_to_be16(MLX4_CQE_STATUS_IPV6)
709 	};
710 
711 	status &= status_all;
712 	switch (status) {
713 	case status_ipv4_tcp:
714 		return (M_HASHTYPE_RSS_TCP_IPV4);
715 	case status_ipv6_tcp:
716 		return (M_HASHTYPE_RSS_TCP_IPV6);
717 	case status_ipv4_udp:
718 		return (udp_rss ? M_HASHTYPE_RSS_UDP_IPV4
719 		    : M_HASHTYPE_RSS_IPV4);
720 	case status_ipv6_udp:
721 		return (udp_rss ? M_HASHTYPE_RSS_UDP_IPV6
722 		    : M_HASHTYPE_RSS_IPV6);
723 	default:
724 		if (status & status_ipv4)
725 			return (M_HASHTYPE_RSS_IPV4);
726 		if (status & status_ipv6)
727 			return (M_HASHTYPE_RSS_IPV6);
728 		return (M_HASHTYPE_OPAQUE_HASH);
729 	}
730 }
731 
732 /* For cpu arch with cache line of 64B the performance is better when cqe size==64B
733  * To enlarge cqe size from 32B to 64B --> 32B of garbage (i.e. 0xccccccc)
734  * was added in the beginning of each cqe (the real data is in the corresponding 32B).
735  * The following calc ensures that when factor==1, it means we are aligned to 64B
736  * and we get the real cqe data*/
737 #define CQE_FACTOR_INDEX(index, factor) (((index) << (factor)) + (factor))
738 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
739 {
740 	struct mlx4_en_priv *priv = netdev_priv(dev);
741 	struct mlx4_cqe *cqe;
742 	struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
743 	struct mlx4_en_rx_mbuf *mb_list;
744 	struct mlx4_en_rx_desc *rx_desc;
745 	struct mbuf *mb;
746 	struct mlx4_cq *mcq = &cq->mcq;
747 	struct mlx4_cqe *buf = cq->buf;
748 	int index;
749 	unsigned int length;
750 	int polled = 0;
751 	u32 cons_index = mcq->cons_index;
752 	u32 size_mask = ring->size_mask;
753 	int size = cq->size;
754 	int factor = priv->cqe_factor;
755 	const int udp_rss = priv->mdev->profile.udp_rss;
756 
757 	if (!priv->port_up)
758 		return 0;
759 
760 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
761 	 * descriptor offset can be deducted from the CQE index instead of
762 	 * reading 'cqe->index' */
763 	index = cons_index & size_mask;
764 	cqe = &buf[CQE_FACTOR_INDEX(index, factor)];
765 
766 	/* Process all completed CQEs */
767 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
768 		    cons_index & size)) {
769 		mb_list = ring->mbuf + index;
770 		rx_desc = ((struct mlx4_en_rx_desc *)ring->buf) + index;
771 
772 		/*
773 		 * make sure we read the CQE after we read the ownership bit
774 		 */
775 		rmb();
776 
777 		if (invalid_cqe(priv, cqe)) {
778 			goto next;
779 		}
780 		/*
781 		 * Packet is OK - process it.
782 		 */
783 		length = be32_to_cpu(cqe->byte_cnt);
784 		length -= ring->fcs_del;
785 
786 		mb = mlx4_en_rx_mb(priv, ring, rx_desc, mb_list, length);
787 		if (unlikely(!mb)) {
788 			ring->errors++;
789 			goto next;
790 		}
791 
792 		ring->bytes += length;
793 		ring->packets++;
794 
795 		if (unlikely(priv->validate_loopback)) {
796 			validate_loopback(priv, mb);
797 			goto next;
798 		}
799 
800 		/* forward Toeplitz compatible hash value */
801 		mb->m_pkthdr.flowid = be32_to_cpu(cqe->immed_rss_invalid);
802 		M_HASHTYPE_SET(mb, mlx4_en_rss_hash(cqe->status, udp_rss));
803 		mb->m_pkthdr.rcvif = dev;
804 		if (be32_to_cpu(cqe->vlan_my_qpn) &
805 		    MLX4_CQE_CVLAN_PRESENT_MASK) {
806 			mb->m_pkthdr.ether_vtag = be16_to_cpu(cqe->sl_vid);
807 			mb->m_flags |= M_VLANTAG;
808 		}
809 		if (likely(dev->if_capenable &
810 		    (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) &&
811 		    (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
812 		    (cqe->checksum == cpu_to_be16(0xffff))) {
813 			priv->port_stats.rx_chksum_good++;
814 			mb->m_pkthdr.csum_flags =
815 			    CSUM_IP_CHECKED | CSUM_IP_VALID |
816 			    CSUM_DATA_VALID | CSUM_PSEUDO_HDR;
817 			mb->m_pkthdr.csum_data = htons(0xffff);
818 			/* This packet is eligible for LRO if it is:
819 			 * - DIX Ethernet (type interpretation)
820 			 * - TCP/IP (v4)
821 			 * - without IP options
822 			 * - not an IP fragment
823 			 */
824 #ifdef INET
825 			if (mlx4_en_can_lro(cqe->status) &&
826 					(dev->if_capenable & IFCAP_LRO)) {
827 				if (ring->lro.lro_cnt != 0 &&
828 						tcp_lro_rx(&ring->lro, mb, 0) == 0)
829 					goto next;
830 			}
831 
832 #endif
833 			/* LRO not possible, complete processing here */
834 			INC_PERF_COUNTER(priv->pstats.lro_misses);
835 		} else {
836 			mb->m_pkthdr.csum_flags = 0;
837 			priv->port_stats.rx_chksum_none++;
838 		}
839 
840 		/* Push it up the stack */
841 		dev->if_input(dev, mb);
842 
843 next:
844 		++cons_index;
845 		index = cons_index & size_mask;
846 		cqe = &buf[CQE_FACTOR_INDEX(index, factor)];
847 		if (++polled == budget)
848 			goto out;
849 	}
850 	/* Flush all pending IP reassembly sessions */
851 out:
852 #ifdef INET
853 	tcp_lro_flush_all(&ring->lro);
854 #endif
855 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
856 	mcq->cons_index = cons_index;
857 	mlx4_cq_set_ci(mcq);
858 	wmb(); /* ensure HW sees CQ consumer before we post new buffers */
859 	ring->cons = mcq->cons_index;
860 	ring->prod += polled; /* Polled descriptors were realocated in place */
861 	mlx4_en_update_rx_prod_db(ring);
862 	return polled;
863 
864 }
865 
866 /* Rx CQ polling - called by NAPI */
867 static int mlx4_en_poll_rx_cq(struct mlx4_en_cq *cq, int budget)
868 {
869 	struct net_device *dev = cq->dev;
870 	struct epoch_tracker et;
871 	int done;
872 
873 	NET_EPOCH_ENTER(et);
874 	done = mlx4_en_process_rx_cq(dev, cq, budget);
875 	NET_EPOCH_EXIT(et);
876 	cq->tot_rx += done;
877 
878 	return done;
879 }
880 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
881 {
882 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
883 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
884         int done;
885 
886         // Shoot one within the irq context
887         // Because there is no NAPI in freeBSD
888         done = mlx4_en_poll_rx_cq(cq, MLX4_EN_RX_BUDGET);
889 	if (priv->port_up  && (done == MLX4_EN_RX_BUDGET) ) {
890 		cq->curr_poll_rx_cpu_id = curcpu;
891 		taskqueue_enqueue(cq->tq, &cq->cq_task);
892         }
893 	else {
894 		mlx4_en_arm_cq(priv, cq);
895 	}
896 }
897 
898 void mlx4_en_rx_que(void *context, int pending)
899 {
900 	struct epoch_tracker et;
901         struct mlx4_en_cq *cq;
902 	struct thread *td;
903 
904         cq = context;
905 	td = curthread;
906 
907 	thread_lock(td);
908 	sched_bind(td, cq->curr_poll_rx_cpu_id);
909 	thread_unlock(td);
910 
911 	NET_EPOCH_ENTER(et);
912         while (mlx4_en_poll_rx_cq(cq, MLX4_EN_RX_BUDGET)
913                         == MLX4_EN_RX_BUDGET);
914 	NET_EPOCH_EXIT(et);
915         mlx4_en_arm_cq(cq->dev->if_softc, cq);
916 }
917 
918 
919 /* RSS related functions */
920 
921 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
922 				 struct mlx4_en_rx_ring *ring,
923 				 enum mlx4_qp_state *state,
924 				 struct mlx4_qp *qp)
925 {
926 	struct mlx4_en_dev *mdev = priv->mdev;
927 	struct mlx4_qp_context *context;
928 	int err = 0;
929 
930 	context = kmalloc(sizeof *context , GFP_KERNEL);
931 	if (!context) {
932 		en_err(priv, "Failed to allocate qp context\n");
933 		return -ENOMEM;
934 	}
935 
936 	err = mlx4_qp_alloc(mdev->dev, qpn, qp, GFP_KERNEL);
937 	if (err) {
938 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
939 		goto out;
940 	}
941 	qp->event = mlx4_en_sqp_event;
942 
943 	memset(context, 0, sizeof *context);
944 	mlx4_en_fill_qp_context(priv, ring->actual_size, sizeof(struct mlx4_en_rx_desc), 0, 0,
945 				qpn, ring->cqn, -1, context);
946 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
947 
948 	/* Cancel FCS removal if FW allows */
949 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
950 		context->param3 |= cpu_to_be32(1 << 29);
951 		ring->fcs_del = ETH_FCS_LEN;
952 	} else
953 		ring->fcs_del = 0;
954 
955 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
956 	if (err) {
957 		mlx4_qp_remove(mdev->dev, qp);
958 		mlx4_qp_free(mdev->dev, qp);
959 	}
960 	mlx4_en_update_rx_prod_db(ring);
961 out:
962 	kfree(context);
963 	return err;
964 }
965 
966 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
967 {
968 	int err;
969 	u32 qpn;
970 
971 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn, 0);
972 	if (err) {
973 		en_err(priv, "Failed reserving drop qpn\n");
974 		return err;
975 	}
976 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp, GFP_KERNEL);
977 	if (err) {
978 		en_err(priv, "Failed allocating drop qp\n");
979 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
980 		return err;
981 	}
982 
983 	return 0;
984 }
985 
986 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
987 {
988 	u32 qpn;
989 
990 	qpn = priv->drop_qp.qpn;
991 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
992 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
993 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
994 }
995 
996 const u32 *
997 mlx4_en_get_rss_key(struct mlx4_en_priv *priv __unused,
998     u16 *keylen)
999 {
1000 	static const u32 rsskey[10] = {
1001 		cpu_to_be32(0xD181C62C),
1002 		cpu_to_be32(0xF7F4DB5B),
1003 		cpu_to_be32(0x1983A2FC),
1004 		cpu_to_be32(0x943E1ADB),
1005 		cpu_to_be32(0xD9389E6B),
1006 		cpu_to_be32(0xD1039C2C),
1007 		cpu_to_be32(0xA74499AD),
1008 		cpu_to_be32(0x593D56D9),
1009 		cpu_to_be32(0xF3253C06),
1010 		cpu_to_be32(0x2ADC1FFC)
1011 	};
1012 
1013 	if (keylen != NULL)
1014 		*keylen = sizeof(rsskey);
1015 	return (rsskey);
1016 }
1017 
1018 u8 mlx4_en_get_rss_mask(struct mlx4_en_priv *priv)
1019 {
1020 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1021 			MLX4_RSS_TCP_IPV6);
1022 
1023 	if (priv->mdev->profile.udp_rss)
1024 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1025 	return (rss_mask);
1026 }
1027 
1028 /* Allocate rx qp's and configure them according to rss map */
1029 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1030 {
1031 	struct mlx4_en_dev *mdev = priv->mdev;
1032 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1033 	struct mlx4_qp_context context;
1034 	struct mlx4_rss_context *rss_context;
1035 	const u32 *key;
1036 	int rss_rings;
1037 	void *ptr;
1038 	int i;
1039 	int err = 0;
1040 	int good_qps = 0;
1041 
1042 	en_dbg(DRV, priv, "Configuring rss steering\n");
1043 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1044 				    priv->rx_ring_num,
1045 				    &rss_map->base_qpn, 0);
1046 	if (err) {
1047 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1048 		return err;
1049 	}
1050 
1051 	for (i = 0; i < priv->rx_ring_num; i++) {
1052 		priv->rx_ring[i]->qpn = rss_map->base_qpn + i;
1053 		err = mlx4_en_config_rss_qp(priv, priv->rx_ring[i]->qpn,
1054 					    priv->rx_ring[i],
1055 					    &rss_map->state[i],
1056 					    &rss_map->qps[i]);
1057 		if (err)
1058 			goto rss_err;
1059 
1060 		++good_qps;
1061 	}
1062 
1063 	/* Configure RSS indirection qp */
1064 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp, GFP_KERNEL);
1065 	if (err) {
1066 		en_err(priv, "Failed to allocate RSS indirection QP\n");
1067 		goto rss_err;
1068 	}
1069 	rss_map->indir_qp.event = mlx4_en_sqp_event;
1070 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1071 				priv->rx_ring[0]->cqn, -1, &context);
1072 
1073 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1074 		rss_rings = priv->rx_ring_num;
1075 	else
1076 		rss_rings = priv->prof->rss_rings;
1077 
1078 	ptr = ((u8 *)&context) + offsetof(struct mlx4_qp_context, pri_path) +
1079 	    MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1080 	rss_context = ptr;
1081 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1082 					    (rss_map->base_qpn));
1083 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1084 	if (priv->mdev->profile.udp_rss)
1085 		rss_context->base_qpn_udp = rss_context->default_qpn;
1086 	rss_context->flags = mlx4_en_get_rss_mask(priv);
1087 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1088 	key = mlx4_en_get_rss_key(priv, NULL);
1089 	for (i = 0; i < 10; i++)
1090 		rss_context->rss_key[i] = key[i];
1091 
1092 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1093 			       &rss_map->indir_qp, &rss_map->indir_state);
1094 	if (err)
1095 		goto indir_err;
1096 
1097 	return 0;
1098 
1099 indir_err:
1100 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1101 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1102 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1103 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1104 rss_err:
1105 	for (i = 0; i < good_qps; i++) {
1106 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1107 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1108 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1109 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1110 	}
1111 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1112 	return err;
1113 }
1114 
1115 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1116 {
1117 	struct mlx4_en_dev *mdev = priv->mdev;
1118 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1119 	int i;
1120 
1121 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1122 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1123 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1124 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1125 
1126 	for (i = 0; i < priv->rx_ring_num; i++) {
1127 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1128 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1129 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1130 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1131 	}
1132 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1133 }
1134 
1135