1 /*
2  * Copyright (c) 2007 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 
34 #include <net/busy_poll.h>
35 #include <linux/mlx4/cq.h>
36 #include <linux/slab.h>
37 #include <linux/mlx4/qp.h>
38 #include <linux/skbuff.h>
39 #include <linux/rculist.h>
40 #include <linux/if_ether.h>
41 #include <linux/if_vlan.h>
42 #include <linux/vmalloc.h>
43 
44 #include "mlx4_en.h"
45 
46 static int mlx4_alloc_pages(struct mlx4_en_priv *priv,
47 			    struct mlx4_en_rx_alloc *page_alloc,
48 			    const struct mlx4_en_frag_info *frag_info,
49 			    gfp_t _gfp)
50 {
51 	int order;
52 	struct page *page;
53 	dma_addr_t dma;
54 
55 	for (order = MLX4_EN_ALLOC_PREFER_ORDER; ;) {
56 		gfp_t gfp = _gfp;
57 
58 		if (order)
59 			gfp |= __GFP_COMP | __GFP_NOWARN;
60 		page = alloc_pages(gfp, order);
61 		if (likely(page))
62 			break;
63 		if (--order < 0 ||
64 		    ((PAGE_SIZE << order) < frag_info->frag_size))
65 			return -ENOMEM;
66 	}
67 	dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE << order,
68 			   PCI_DMA_FROMDEVICE);
69 	if (dma_mapping_error(priv->ddev, dma)) {
70 		put_page(page);
71 		return -ENOMEM;
72 	}
73 	page_alloc->page_size = PAGE_SIZE << order;
74 	page_alloc->page = page;
75 	page_alloc->dma = dma;
76 	page_alloc->page_offset = frag_info->frag_align;
77 	/* Not doing get_page() for each frag is a big win
78 	 * on asymetric workloads.
79 	 */
80 	atomic_set(&page->_count,
81 		   page_alloc->page_size / frag_info->frag_stride);
82 	return 0;
83 }
84 
85 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
86 			       struct mlx4_en_rx_desc *rx_desc,
87 			       struct mlx4_en_rx_alloc *frags,
88 			       struct mlx4_en_rx_alloc *ring_alloc,
89 			       gfp_t gfp)
90 {
91 	struct mlx4_en_rx_alloc page_alloc[MLX4_EN_MAX_RX_FRAGS];
92 	const struct mlx4_en_frag_info *frag_info;
93 	struct page *page;
94 	dma_addr_t dma;
95 	int i;
96 
97 	for (i = 0; i < priv->num_frags; i++) {
98 		frag_info = &priv->frag_info[i];
99 		page_alloc[i] = ring_alloc[i];
100 		page_alloc[i].page_offset += frag_info->frag_stride;
101 
102 		if (page_alloc[i].page_offset + frag_info->frag_stride <=
103 		    ring_alloc[i].page_size)
104 			continue;
105 
106 		if (mlx4_alloc_pages(priv, &page_alloc[i], frag_info, gfp))
107 			goto out;
108 	}
109 
110 	for (i = 0; i < priv->num_frags; i++) {
111 		frags[i] = ring_alloc[i];
112 		dma = ring_alloc[i].dma + ring_alloc[i].page_offset;
113 		ring_alloc[i] = page_alloc[i];
114 		rx_desc->data[i].addr = cpu_to_be64(dma);
115 	}
116 
117 	return 0;
118 
119 out:
120 	while (i--) {
121 		frag_info = &priv->frag_info[i];
122 		if (page_alloc[i].page != ring_alloc[i].page) {
123 			dma_unmap_page(priv->ddev, page_alloc[i].dma,
124 				page_alloc[i].page_size, PCI_DMA_FROMDEVICE);
125 			page = page_alloc[i].page;
126 			atomic_set(&page->_count, 1);
127 			put_page(page);
128 		}
129 	}
130 	return -ENOMEM;
131 }
132 
133 static void mlx4_en_free_frag(struct mlx4_en_priv *priv,
134 			      struct mlx4_en_rx_alloc *frags,
135 			      int i)
136 {
137 	const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
138 	u32 next_frag_end = frags[i].page_offset + 2 * frag_info->frag_stride;
139 
140 
141 	if (next_frag_end > frags[i].page_size)
142 		dma_unmap_page(priv->ddev, frags[i].dma, frags[i].page_size,
143 			       PCI_DMA_FROMDEVICE);
144 
145 	if (frags[i].page)
146 		put_page(frags[i].page);
147 }
148 
149 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
150 				  struct mlx4_en_rx_ring *ring)
151 {
152 	int i;
153 	struct mlx4_en_rx_alloc *page_alloc;
154 
155 	for (i = 0; i < priv->num_frags; i++) {
156 		const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
157 
158 		if (mlx4_alloc_pages(priv, &ring->page_alloc[i],
159 				     frag_info, GFP_KERNEL))
160 			goto out;
161 	}
162 	return 0;
163 
164 out:
165 	while (i--) {
166 		struct page *page;
167 
168 		page_alloc = &ring->page_alloc[i];
169 		dma_unmap_page(priv->ddev, page_alloc->dma,
170 			       page_alloc->page_size, PCI_DMA_FROMDEVICE);
171 		page = page_alloc->page;
172 		atomic_set(&page->_count, 1);
173 		put_page(page);
174 		page_alloc->page = NULL;
175 	}
176 	return -ENOMEM;
177 }
178 
179 static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
180 				      struct mlx4_en_rx_ring *ring)
181 {
182 	struct mlx4_en_rx_alloc *page_alloc;
183 	int i;
184 
185 	for (i = 0; i < priv->num_frags; i++) {
186 		const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
187 
188 		page_alloc = &ring->page_alloc[i];
189 		en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
190 		       i, page_count(page_alloc->page));
191 
192 		dma_unmap_page(priv->ddev, page_alloc->dma,
193 				page_alloc->page_size, PCI_DMA_FROMDEVICE);
194 		while (page_alloc->page_offset + frag_info->frag_stride <
195 		       page_alloc->page_size) {
196 			put_page(page_alloc->page);
197 			page_alloc->page_offset += frag_info->frag_stride;
198 		}
199 		page_alloc->page = NULL;
200 	}
201 }
202 
203 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
204 				 struct mlx4_en_rx_ring *ring, int index)
205 {
206 	struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
207 	int possible_frags;
208 	int i;
209 
210 	/* Set size and memtype fields */
211 	for (i = 0; i < priv->num_frags; i++) {
212 		rx_desc->data[i].byte_count =
213 			cpu_to_be32(priv->frag_info[i].frag_size);
214 		rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
215 	}
216 
217 	/* If the number of used fragments does not fill up the ring stride,
218 	 * remaining (unused) fragments must be padded with null address/size
219 	 * and a special memory key */
220 	possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
221 	for (i = priv->num_frags; i < possible_frags; i++) {
222 		rx_desc->data[i].byte_count = 0;
223 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
224 		rx_desc->data[i].addr = 0;
225 	}
226 }
227 
228 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
229 				   struct mlx4_en_rx_ring *ring, int index,
230 				   gfp_t gfp)
231 {
232 	struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
233 	struct mlx4_en_rx_alloc *frags = ring->rx_info +
234 					(index << priv->log_rx_info);
235 
236 	return mlx4_en_alloc_frags(priv, rx_desc, frags, ring->page_alloc, gfp);
237 }
238 
239 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
240 {
241 	*ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
242 }
243 
244 static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
245 				 struct mlx4_en_rx_ring *ring,
246 				 int index)
247 {
248 	struct mlx4_en_rx_alloc *frags;
249 	int nr;
250 
251 	frags = ring->rx_info + (index << priv->log_rx_info);
252 	for (nr = 0; nr < priv->num_frags; nr++) {
253 		en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
254 		mlx4_en_free_frag(priv, frags, nr);
255 	}
256 }
257 
258 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
259 {
260 	struct mlx4_en_rx_ring *ring;
261 	int ring_ind;
262 	int buf_ind;
263 	int new_size;
264 
265 	for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
266 		for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
267 			ring = priv->rx_ring[ring_ind];
268 
269 			if (mlx4_en_prepare_rx_desc(priv, ring,
270 						    ring->actual_size,
271 						    GFP_KERNEL)) {
272 				if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
273 					en_err(priv, "Failed to allocate "
274 						     "enough rx buffers\n");
275 					return -ENOMEM;
276 				} else {
277 					new_size = rounddown_pow_of_two(ring->actual_size);
278 					en_warn(priv, "Only %d buffers allocated "
279 						      "reducing ring size to %d",
280 						ring->actual_size, new_size);
281 					goto reduce_rings;
282 				}
283 			}
284 			ring->actual_size++;
285 			ring->prod++;
286 		}
287 	}
288 	return 0;
289 
290 reduce_rings:
291 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
292 		ring = priv->rx_ring[ring_ind];
293 		while (ring->actual_size > new_size) {
294 			ring->actual_size--;
295 			ring->prod--;
296 			mlx4_en_free_rx_desc(priv, ring, 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_rx_desc(priv, ring, index);
317 		++ring->cons;
318 	}
319 }
320 
321 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
322 			   struct mlx4_en_rx_ring **pring,
323 			   u32 size, u16 stride, int node)
324 {
325 	struct mlx4_en_dev *mdev = priv->mdev;
326 	struct mlx4_en_rx_ring *ring;
327 	int err = -ENOMEM;
328 	int tmp;
329 
330 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
331 	if (!ring) {
332 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
333 		if (!ring) {
334 			en_err(priv, "Failed to allocate RX ring structure\n");
335 			return -ENOMEM;
336 		}
337 	}
338 
339 	ring->prod = 0;
340 	ring->cons = 0;
341 	ring->size = size;
342 	ring->size_mask = size - 1;
343 	ring->stride = stride;
344 	ring->log_stride = ffs(ring->stride) - 1;
345 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
346 
347 	tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
348 					sizeof(struct mlx4_en_rx_alloc));
349 	ring->rx_info = vmalloc_node(tmp, node);
350 	if (!ring->rx_info) {
351 		ring->rx_info = vmalloc(tmp);
352 		if (!ring->rx_info) {
353 			err = -ENOMEM;
354 			goto err_ring;
355 		}
356 	}
357 
358 	en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
359 		 ring->rx_info, tmp);
360 
361 	/* Allocate HW buffers on provided NUMA node */
362 	set_dev_node(&mdev->dev->pdev->dev, node);
363 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
364 				 ring->buf_size, 2 * PAGE_SIZE);
365 	set_dev_node(&mdev->dev->pdev->dev, mdev->dev->numa_node);
366 	if (err)
367 		goto err_info;
368 
369 	err = mlx4_en_map_buffer(&ring->wqres.buf);
370 	if (err) {
371 		en_err(priv, "Failed to map RX buffer\n");
372 		goto err_hwq;
373 	}
374 	ring->buf = ring->wqres.buf.direct.buf;
375 
376 	ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
377 
378 	*pring = ring;
379 	return 0;
380 
381 err_hwq:
382 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
383 err_info:
384 	vfree(ring->rx_info);
385 	ring->rx_info = NULL;
386 err_ring:
387 	kfree(ring);
388 	*pring = NULL;
389 
390 	return err;
391 }
392 
393 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
394 {
395 	struct mlx4_en_rx_ring *ring;
396 	int i;
397 	int ring_ind;
398 	int err;
399 	int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
400 					DS_SIZE * priv->num_frags);
401 
402 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
403 		ring = priv->rx_ring[ring_ind];
404 
405 		ring->prod = 0;
406 		ring->cons = 0;
407 		ring->actual_size = 0;
408 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
409 
410 		ring->stride = stride;
411 		if (ring->stride <= TXBB_SIZE)
412 			ring->buf += TXBB_SIZE;
413 
414 		ring->log_stride = ffs(ring->stride) - 1;
415 		ring->buf_size = ring->size * ring->stride;
416 
417 		memset(ring->buf, 0, ring->buf_size);
418 		mlx4_en_update_rx_prod_db(ring);
419 
420 		/* Initialize all descriptors */
421 		for (i = 0; i < ring->size; i++)
422 			mlx4_en_init_rx_desc(priv, ring, i);
423 
424 		/* Initialize page allocators */
425 		err = mlx4_en_init_allocator(priv, ring);
426 		if (err) {
427 			en_err(priv, "Failed initializing ring allocator\n");
428 			if (ring->stride <= TXBB_SIZE)
429 				ring->buf -= TXBB_SIZE;
430 			ring_ind--;
431 			goto err_allocator;
432 		}
433 	}
434 	err = mlx4_en_fill_rx_buffers(priv);
435 	if (err)
436 		goto err_buffers;
437 
438 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
439 		ring = priv->rx_ring[ring_ind];
440 
441 		ring->size_mask = ring->actual_size - 1;
442 		mlx4_en_update_rx_prod_db(ring);
443 	}
444 
445 	return 0;
446 
447 err_buffers:
448 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
449 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
450 
451 	ring_ind = priv->rx_ring_num - 1;
452 err_allocator:
453 	while (ring_ind >= 0) {
454 		if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
455 			priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
456 		mlx4_en_destroy_allocator(priv, priv->rx_ring[ring_ind]);
457 		ring_ind--;
458 	}
459 	return err;
460 }
461 
462 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
463 			     struct mlx4_en_rx_ring **pring,
464 			     u32 size, u16 stride)
465 {
466 	struct mlx4_en_dev *mdev = priv->mdev;
467 	struct mlx4_en_rx_ring *ring = *pring;
468 
469 	mlx4_en_unmap_buffer(&ring->wqres.buf);
470 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
471 	vfree(ring->rx_info);
472 	ring->rx_info = NULL;
473 	kfree(ring);
474 	*pring = NULL;
475 #ifdef CONFIG_RFS_ACCEL
476 	mlx4_en_cleanup_filters(priv);
477 #endif
478 }
479 
480 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
481 				struct mlx4_en_rx_ring *ring)
482 {
483 	mlx4_en_free_rx_buf(priv, ring);
484 	if (ring->stride <= TXBB_SIZE)
485 		ring->buf -= TXBB_SIZE;
486 	mlx4_en_destroy_allocator(priv, ring);
487 }
488 
489 
490 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
491 				    struct mlx4_en_rx_desc *rx_desc,
492 				    struct mlx4_en_rx_alloc *frags,
493 				    struct sk_buff *skb,
494 				    int length)
495 {
496 	struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
497 	struct mlx4_en_frag_info *frag_info;
498 	int nr;
499 	dma_addr_t dma;
500 
501 	/* Collect used fragments while replacing them in the HW descriptors */
502 	for (nr = 0; nr < priv->num_frags; nr++) {
503 		frag_info = &priv->frag_info[nr];
504 		if (length <= frag_info->frag_prefix_size)
505 			break;
506 		if (!frags[nr].page)
507 			goto fail;
508 
509 		dma = be64_to_cpu(rx_desc->data[nr].addr);
510 		dma_sync_single_for_cpu(priv->ddev, dma, frag_info->frag_size,
511 					DMA_FROM_DEVICE);
512 
513 		/* Save page reference in skb */
514 		__skb_frag_set_page(&skb_frags_rx[nr], frags[nr].page);
515 		skb_frag_size_set(&skb_frags_rx[nr], frag_info->frag_size);
516 		skb_frags_rx[nr].page_offset = frags[nr].page_offset;
517 		skb->truesize += frag_info->frag_stride;
518 		frags[nr].page = NULL;
519 	}
520 	/* Adjust size of last fragment to match actual length */
521 	if (nr > 0)
522 		skb_frag_size_set(&skb_frags_rx[nr - 1],
523 			length - priv->frag_info[nr - 1].frag_prefix_size);
524 	return nr;
525 
526 fail:
527 	while (nr > 0) {
528 		nr--;
529 		__skb_frag_unref(&skb_frags_rx[nr]);
530 	}
531 	return 0;
532 }
533 
534 
535 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
536 				      struct mlx4_en_rx_desc *rx_desc,
537 				      struct mlx4_en_rx_alloc *frags,
538 				      unsigned int length)
539 {
540 	struct sk_buff *skb;
541 	void *va;
542 	int used_frags;
543 	dma_addr_t dma;
544 
545 	skb = netdev_alloc_skb(priv->dev, SMALL_PACKET_SIZE + NET_IP_ALIGN);
546 	if (!skb) {
547 		en_dbg(RX_ERR, priv, "Failed allocating skb\n");
548 		return NULL;
549 	}
550 	skb_reserve(skb, NET_IP_ALIGN);
551 	skb->len = length;
552 
553 	/* Get pointer to first fragment so we could copy the headers into the
554 	 * (linear part of the) skb */
555 	va = page_address(frags[0].page) + frags[0].page_offset;
556 
557 	if (length <= SMALL_PACKET_SIZE) {
558 		/* We are copying all relevant data to the skb - temporarily
559 		 * sync buffers for the copy */
560 		dma = be64_to_cpu(rx_desc->data[0].addr);
561 		dma_sync_single_for_cpu(priv->ddev, dma, length,
562 					DMA_FROM_DEVICE);
563 		skb_copy_to_linear_data(skb, va, length);
564 		skb->tail += length;
565 	} else {
566 		/* Move relevant fragments to skb */
567 		used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, frags,
568 							skb, length);
569 		if (unlikely(!used_frags)) {
570 			kfree_skb(skb);
571 			return NULL;
572 		}
573 		skb_shinfo(skb)->nr_frags = used_frags;
574 
575 		/* Copy headers into the skb linear buffer */
576 		memcpy(skb->data, va, HEADER_COPY_SIZE);
577 		skb->tail += HEADER_COPY_SIZE;
578 
579 		/* Skip headers in first fragment */
580 		skb_shinfo(skb)->frags[0].page_offset += HEADER_COPY_SIZE;
581 
582 		/* Adjust size of first fragment */
583 		skb_frag_size_sub(&skb_shinfo(skb)->frags[0], HEADER_COPY_SIZE);
584 		skb->data_len = length - HEADER_COPY_SIZE;
585 	}
586 	return skb;
587 }
588 
589 static void validate_loopback(struct mlx4_en_priv *priv, struct sk_buff *skb)
590 {
591 	int i;
592 	int offset = ETH_HLEN;
593 
594 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
595 		if (*(skb->data + offset) != (unsigned char) (i & 0xff))
596 			goto out_loopback;
597 	}
598 	/* Loopback found */
599 	priv->loopback_ok = 1;
600 
601 out_loopback:
602 	dev_kfree_skb_any(skb);
603 }
604 
605 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
606 				     struct mlx4_en_rx_ring *ring)
607 {
608 	int index = ring->prod & ring->size_mask;
609 
610 	while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
611 		if (mlx4_en_prepare_rx_desc(priv, ring, index, GFP_ATOMIC))
612 			break;
613 		ring->prod++;
614 		index = ring->prod & ring->size_mask;
615 	}
616 }
617 
618 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
619 {
620 	struct mlx4_en_priv *priv = netdev_priv(dev);
621 	struct mlx4_en_dev *mdev = priv->mdev;
622 	struct mlx4_cqe *cqe;
623 	struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
624 	struct mlx4_en_rx_alloc *frags;
625 	struct mlx4_en_rx_desc *rx_desc;
626 	struct sk_buff *skb;
627 	int index;
628 	int nr;
629 	unsigned int length;
630 	int polled = 0;
631 	int ip_summed;
632 	int factor = priv->cqe_factor;
633 	u64 timestamp;
634 	bool l2_tunnel;
635 
636 	if (!priv->port_up)
637 		return 0;
638 
639 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
640 	 * descriptor offset can be deduced from the CQE index instead of
641 	 * reading 'cqe->index' */
642 	index = cq->mcq.cons_index & ring->size_mask;
643 	cqe = &cq->buf[(index << factor) + factor];
644 
645 	/* Process all completed CQEs */
646 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
647 		    cq->mcq.cons_index & cq->size)) {
648 
649 		frags = ring->rx_info + (index << priv->log_rx_info);
650 		rx_desc = ring->buf + (index << ring->log_stride);
651 
652 		/*
653 		 * make sure we read the CQE after we read the ownership bit
654 		 */
655 		rmb();
656 
657 		/* Drop packet on bad receive or bad checksum */
658 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
659 						MLX4_CQE_OPCODE_ERROR)) {
660 			en_err(priv, "CQE completed in error - vendor "
661 				  "syndrom:%d syndrom:%d\n",
662 				  ((struct mlx4_err_cqe *) cqe)->vendor_err_syndrome,
663 				  ((struct mlx4_err_cqe *) cqe)->syndrome);
664 			goto next;
665 		}
666 		if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
667 			en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
668 			goto next;
669 		}
670 
671 		/* Check if we need to drop the packet if SRIOV is not enabled
672 		 * and not performing the selftest or flb disabled
673 		 */
674 		if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
675 			struct ethhdr *ethh;
676 			dma_addr_t dma;
677 			/* Get pointer to first fragment since we haven't
678 			 * skb yet and cast it to ethhdr struct
679 			 */
680 			dma = be64_to_cpu(rx_desc->data[0].addr);
681 			dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
682 						DMA_FROM_DEVICE);
683 			ethh = (struct ethhdr *)(page_address(frags[0].page) +
684 						 frags[0].page_offset);
685 
686 			if (is_multicast_ether_addr(ethh->h_dest)) {
687 				struct mlx4_mac_entry *entry;
688 				struct hlist_head *bucket;
689 				unsigned int mac_hash;
690 
691 				/* Drop the packet, since HW loopback-ed it */
692 				mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
693 				bucket = &priv->mac_hash[mac_hash];
694 				rcu_read_lock();
695 				hlist_for_each_entry_rcu(entry, bucket, hlist) {
696 					if (ether_addr_equal_64bits(entry->mac,
697 								    ethh->h_source)) {
698 						rcu_read_unlock();
699 						goto next;
700 					}
701 				}
702 				rcu_read_unlock();
703 			}
704 		}
705 
706 		/*
707 		 * Packet is OK - process it.
708 		 */
709 		length = be32_to_cpu(cqe->byte_cnt);
710 		length -= ring->fcs_del;
711 		ring->bytes += length;
712 		ring->packets++;
713 		l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
714 			(cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
715 
716 		if (likely(dev->features & NETIF_F_RXCSUM)) {
717 			if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
718 			    (cqe->checksum == cpu_to_be16(0xffff))) {
719 				ring->csum_ok++;
720 				/* This packet is eligible for GRO if it is:
721 				 * - DIX Ethernet (type interpretation)
722 				 * - TCP/IP (v4)
723 				 * - without IP options
724 				 * - not an IP fragment
725 				 * - no LLS polling in progress
726 				 */
727 				if (!mlx4_en_cq_busy_polling(cq) &&
728 				    (dev->features & NETIF_F_GRO)) {
729 					struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
730 					if (!gro_skb)
731 						goto next;
732 
733 					nr = mlx4_en_complete_rx_desc(priv,
734 						rx_desc, frags, gro_skb,
735 						length);
736 					if (!nr)
737 						goto next;
738 
739 					skb_shinfo(gro_skb)->nr_frags = nr;
740 					gro_skb->len = length;
741 					gro_skb->data_len = length;
742 					gro_skb->ip_summed = CHECKSUM_UNNECESSARY;
743 
744 					if (l2_tunnel)
745 						gro_skb->encapsulation = 1;
746 					if ((cqe->vlan_my_qpn &
747 					    cpu_to_be32(MLX4_CQE_VLAN_PRESENT_MASK)) &&
748 					    (dev->features & NETIF_F_HW_VLAN_CTAG_RX)) {
749 						u16 vid = be16_to_cpu(cqe->sl_vid);
750 
751 						__vlan_hwaccel_put_tag(gro_skb, htons(ETH_P_8021Q), vid);
752 					}
753 
754 					if (dev->features & NETIF_F_RXHASH)
755 						skb_set_hash(gro_skb,
756 							     be32_to_cpu(cqe->immed_rss_invalid),
757 							     PKT_HASH_TYPE_L3);
758 
759 					skb_record_rx_queue(gro_skb, cq->ring);
760 
761 					if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
762 						timestamp = mlx4_en_get_cqe_ts(cqe);
763 						mlx4_en_fill_hwtstamps(mdev,
764 								       skb_hwtstamps(gro_skb),
765 								       timestamp);
766 					}
767 
768 					napi_gro_frags(&cq->napi);
769 					goto next;
770 				}
771 
772 				/* GRO not possible, complete processing here */
773 				ip_summed = CHECKSUM_UNNECESSARY;
774 			} else {
775 				ip_summed = CHECKSUM_NONE;
776 				ring->csum_none++;
777 			}
778 		} else {
779 			ip_summed = CHECKSUM_NONE;
780 			ring->csum_none++;
781 		}
782 
783 		skb = mlx4_en_rx_skb(priv, rx_desc, frags, length);
784 		if (!skb) {
785 			priv->stats.rx_dropped++;
786 			goto next;
787 		}
788 
789                 if (unlikely(priv->validate_loopback)) {
790 			validate_loopback(priv, skb);
791 			goto next;
792 		}
793 
794 		skb->ip_summed = ip_summed;
795 		skb->protocol = eth_type_trans(skb, dev);
796 		skb_record_rx_queue(skb, cq->ring);
797 
798 		if (l2_tunnel)
799 			skb->encapsulation = 1;
800 
801 		if (dev->features & NETIF_F_RXHASH)
802 			skb_set_hash(skb,
803 				     be32_to_cpu(cqe->immed_rss_invalid),
804 				     PKT_HASH_TYPE_L3);
805 
806 		if ((be32_to_cpu(cqe->vlan_my_qpn) &
807 		    MLX4_CQE_VLAN_PRESENT_MASK) &&
808 		    (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
809 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), be16_to_cpu(cqe->sl_vid));
810 
811 		if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
812 			timestamp = mlx4_en_get_cqe_ts(cqe);
813 			mlx4_en_fill_hwtstamps(mdev, skb_hwtstamps(skb),
814 					       timestamp);
815 		}
816 
817 		skb_mark_napi_id(skb, &cq->napi);
818 
819 		if (!mlx4_en_cq_busy_polling(cq))
820 			napi_gro_receive(&cq->napi, skb);
821 		else
822 			netif_receive_skb(skb);
823 
824 next:
825 		for (nr = 0; nr < priv->num_frags; nr++)
826 			mlx4_en_free_frag(priv, frags, nr);
827 
828 		++cq->mcq.cons_index;
829 		index = (cq->mcq.cons_index) & ring->size_mask;
830 		cqe = &cq->buf[(index << factor) + factor];
831 		if (++polled == budget)
832 			goto out;
833 	}
834 
835 out:
836 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
837 	mlx4_cq_set_ci(&cq->mcq);
838 	wmb(); /* ensure HW sees CQ consumer before we post new buffers */
839 	ring->cons = cq->mcq.cons_index;
840 	mlx4_en_refill_rx_buffers(priv, ring);
841 	mlx4_en_update_rx_prod_db(ring);
842 	return polled;
843 }
844 
845 
846 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
847 {
848 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
849 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
850 
851 	if (priv->port_up)
852 		napi_schedule(&cq->napi);
853 	else
854 		mlx4_en_arm_cq(priv, cq);
855 }
856 
857 /* Rx CQ polling - called by NAPI */
858 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
859 {
860 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
861 	struct net_device *dev = cq->dev;
862 	struct mlx4_en_priv *priv = netdev_priv(dev);
863 	int done;
864 
865 	if (!mlx4_en_cq_lock_napi(cq))
866 		return budget;
867 
868 	done = mlx4_en_process_rx_cq(dev, cq, budget);
869 
870 	mlx4_en_cq_unlock_napi(cq);
871 
872 	/* If we used up all the quota - we're probably not done yet... */
873 	if (done == budget)
874 		INC_PERF_COUNTER(priv->pstats.napi_quota);
875 	else {
876 		/* Done for now */
877 		napi_complete(napi);
878 		mlx4_en_arm_cq(priv, cq);
879 	}
880 	return done;
881 }
882 
883 static const int frag_sizes[] = {
884 	FRAG_SZ0,
885 	FRAG_SZ1,
886 	FRAG_SZ2,
887 	FRAG_SZ3
888 };
889 
890 void mlx4_en_calc_rx_buf(struct net_device *dev)
891 {
892 	struct mlx4_en_priv *priv = netdev_priv(dev);
893 	int eff_mtu = dev->mtu + ETH_HLEN + VLAN_HLEN + ETH_LLC_SNAP_SIZE;
894 	int buf_size = 0;
895 	int i = 0;
896 
897 	while (buf_size < eff_mtu) {
898 		priv->frag_info[i].frag_size =
899 			(eff_mtu > buf_size + frag_sizes[i]) ?
900 				frag_sizes[i] : eff_mtu - buf_size;
901 		priv->frag_info[i].frag_prefix_size = buf_size;
902 		if (!i)	{
903 			priv->frag_info[i].frag_align = NET_IP_ALIGN;
904 			priv->frag_info[i].frag_stride =
905 				ALIGN(frag_sizes[i] + NET_IP_ALIGN, SMP_CACHE_BYTES);
906 		} else {
907 			priv->frag_info[i].frag_align = 0;
908 			priv->frag_info[i].frag_stride =
909 				ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
910 		}
911 		buf_size += priv->frag_info[i].frag_size;
912 		i++;
913 	}
914 
915 	priv->num_frags = i;
916 	priv->rx_skb_size = eff_mtu;
917 	priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
918 
919 	en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
920 		  "num_frags:%d):\n", eff_mtu, priv->num_frags);
921 	for (i = 0; i < priv->num_frags; i++) {
922 		en_err(priv,
923 		       "  frag:%d - size:%d prefix:%d align:%d stride:%d\n",
924 		       i,
925 		       priv->frag_info[i].frag_size,
926 		       priv->frag_info[i].frag_prefix_size,
927 		       priv->frag_info[i].frag_align,
928 		       priv->frag_info[i].frag_stride);
929 	}
930 }
931 
932 /* RSS related functions */
933 
934 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
935 				 struct mlx4_en_rx_ring *ring,
936 				 enum mlx4_qp_state *state,
937 				 struct mlx4_qp *qp)
938 {
939 	struct mlx4_en_dev *mdev = priv->mdev;
940 	struct mlx4_qp_context *context;
941 	int err = 0;
942 
943 	context = kmalloc(sizeof(*context), GFP_KERNEL);
944 	if (!context)
945 		return -ENOMEM;
946 
947 	err = mlx4_qp_alloc(mdev->dev, qpn, qp);
948 	if (err) {
949 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
950 		goto out;
951 	}
952 	qp->event = mlx4_en_sqp_event;
953 
954 	memset(context, 0, sizeof *context);
955 	mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
956 				qpn, ring->cqn, -1, context);
957 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
958 
959 	/* Cancel FCS removal if FW allows */
960 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
961 		context->param3 |= cpu_to_be32(1 << 29);
962 		ring->fcs_del = ETH_FCS_LEN;
963 	} else
964 		ring->fcs_del = 0;
965 
966 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
967 	if (err) {
968 		mlx4_qp_remove(mdev->dev, qp);
969 		mlx4_qp_free(mdev->dev, qp);
970 	}
971 	mlx4_en_update_rx_prod_db(ring);
972 out:
973 	kfree(context);
974 	return err;
975 }
976 
977 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
978 {
979 	int err;
980 	u32 qpn;
981 
982 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn);
983 	if (err) {
984 		en_err(priv, "Failed reserving drop qpn\n");
985 		return err;
986 	}
987 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp);
988 	if (err) {
989 		en_err(priv, "Failed allocating drop qp\n");
990 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
991 		return err;
992 	}
993 
994 	return 0;
995 }
996 
997 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
998 {
999 	u32 qpn;
1000 
1001 	qpn = priv->drop_qp.qpn;
1002 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1003 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1004 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1005 }
1006 
1007 /* Allocate rx qp's and configure them according to rss map */
1008 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1009 {
1010 	struct mlx4_en_dev *mdev = priv->mdev;
1011 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1012 	struct mlx4_qp_context context;
1013 	struct mlx4_rss_context *rss_context;
1014 	int rss_rings;
1015 	void *ptr;
1016 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1017 			MLX4_RSS_TCP_IPV6);
1018 	int i, qpn;
1019 	int err = 0;
1020 	int good_qps = 0;
1021 	static const u32 rsskey[10] = { 0xD181C62C, 0xF7F4DB5B, 0x1983A2FC,
1022 				0x943E1ADB, 0xD9389E6B, 0xD1039C2C, 0xA74499AD,
1023 				0x593D56D9, 0xF3253C06, 0x2ADC1FFC};
1024 
1025 	en_dbg(DRV, priv, "Configuring rss steering\n");
1026 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1027 				    priv->rx_ring_num,
1028 				    &rss_map->base_qpn);
1029 	if (err) {
1030 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1031 		return err;
1032 	}
1033 
1034 	for (i = 0; i < priv->rx_ring_num; i++) {
1035 		qpn = rss_map->base_qpn + i;
1036 		err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1037 					    &rss_map->state[i],
1038 					    &rss_map->qps[i]);
1039 		if (err)
1040 			goto rss_err;
1041 
1042 		++good_qps;
1043 	}
1044 
1045 	/* Configure RSS indirection qp */
1046 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
1047 	if (err) {
1048 		en_err(priv, "Failed to allocate RSS indirection QP\n");
1049 		goto rss_err;
1050 	}
1051 	rss_map->indir_qp.event = mlx4_en_sqp_event;
1052 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1053 				priv->rx_ring[0]->cqn, -1, &context);
1054 
1055 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1056 		rss_rings = priv->rx_ring_num;
1057 	else
1058 		rss_rings = priv->prof->rss_rings;
1059 
1060 	ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1061 					+ MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1062 	rss_context = ptr;
1063 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1064 					    (rss_map->base_qpn));
1065 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1066 	if (priv->mdev->profile.udp_rss) {
1067 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1068 		rss_context->base_qpn_udp = rss_context->default_qpn;
1069 	}
1070 
1071 	if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1072 		en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1073 		rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1074 	}
1075 
1076 	rss_context->flags = rss_mask;
1077 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1078 	for (i = 0; i < 10; i++)
1079 		rss_context->rss_key[i] = cpu_to_be32(rsskey[i]);
1080 
1081 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1082 			       &rss_map->indir_qp, &rss_map->indir_state);
1083 	if (err)
1084 		goto indir_err;
1085 
1086 	return 0;
1087 
1088 indir_err:
1089 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1090 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1091 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1092 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1093 rss_err:
1094 	for (i = 0; i < good_qps; i++) {
1095 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1096 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1097 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1098 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1099 	}
1100 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1101 	return err;
1102 }
1103 
1104 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1105 {
1106 	struct mlx4_en_dev *mdev = priv->mdev;
1107 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1108 	int i;
1109 
1110 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1111 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1112 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1113 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1114 
1115 	for (i = 0; i < priv->rx_ring_num; i++) {
1116 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1117 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1118 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1119 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1120 	}
1121 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1122 }
1123