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 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 		if (!dev->caps.comp_pool)
330 			num_of_eqs = max_t(int, MIN_RX_RINGS,
331 					   min_t(int,
332 						 dev->caps.num_comp_vectors,
333 						 DEF_RX_RINGS));
334 		else
335 			num_of_eqs = min_t(int, MAX_MSIX_P_PORT,
336 					   dev->caps.comp_pool/
337 					   dev->caps.num_ports) - 1;
338 
339 		num_rx_rings = min_t(int, num_of_eqs,
340 				     netif_get_num_default_rss_queues());
341 		mdev->profile.prof[i].rx_ring_num =
342 			rounddown_pow_of_two(num_rx_rings);
343 	}
344 }
345 
346 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
347 			   struct mlx4_en_rx_ring **pring,
348 			   u32 size, u16 stride, int node)
349 {
350 	struct mlx4_en_dev *mdev = priv->mdev;
351 	struct mlx4_en_rx_ring *ring;
352 	int err = -ENOMEM;
353 	int tmp;
354 
355 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
356 	if (!ring) {
357 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
358 		if (!ring) {
359 			en_err(priv, "Failed to allocate RX ring structure\n");
360 			return -ENOMEM;
361 		}
362 	}
363 
364 	ring->prod = 0;
365 	ring->cons = 0;
366 	ring->size = size;
367 	ring->size_mask = size - 1;
368 	ring->stride = stride;
369 	ring->log_stride = ffs(ring->stride) - 1;
370 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
371 
372 	tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
373 					sizeof(struct mlx4_en_rx_alloc));
374 	ring->rx_info = vmalloc_node(tmp, node);
375 	if (!ring->rx_info) {
376 		ring->rx_info = vmalloc(tmp);
377 		if (!ring->rx_info) {
378 			err = -ENOMEM;
379 			goto err_ring;
380 		}
381 	}
382 
383 	en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
384 		 ring->rx_info, tmp);
385 
386 	/* Allocate HW buffers on provided NUMA node */
387 	set_dev_node(&mdev->dev->pdev->dev, node);
388 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
389 				 ring->buf_size, 2 * PAGE_SIZE);
390 	set_dev_node(&mdev->dev->pdev->dev, mdev->dev->numa_node);
391 	if (err)
392 		goto err_info;
393 
394 	err = mlx4_en_map_buffer(&ring->wqres.buf);
395 	if (err) {
396 		en_err(priv, "Failed to map RX buffer\n");
397 		goto err_hwq;
398 	}
399 	ring->buf = ring->wqres.buf.direct.buf;
400 
401 	ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
402 
403 	*pring = ring;
404 	return 0;
405 
406 err_hwq:
407 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
408 err_info:
409 	vfree(ring->rx_info);
410 	ring->rx_info = NULL;
411 err_ring:
412 	kfree(ring);
413 	*pring = NULL;
414 
415 	return err;
416 }
417 
418 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
419 {
420 	struct mlx4_en_rx_ring *ring;
421 	int i;
422 	int ring_ind;
423 	int err;
424 	int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
425 					DS_SIZE * priv->num_frags);
426 
427 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
428 		ring = priv->rx_ring[ring_ind];
429 
430 		ring->prod = 0;
431 		ring->cons = 0;
432 		ring->actual_size = 0;
433 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
434 
435 		ring->stride = stride;
436 		if (ring->stride <= TXBB_SIZE)
437 			ring->buf += TXBB_SIZE;
438 
439 		ring->log_stride = ffs(ring->stride) - 1;
440 		ring->buf_size = ring->size * ring->stride;
441 
442 		memset(ring->buf, 0, ring->buf_size);
443 		mlx4_en_update_rx_prod_db(ring);
444 
445 		/* Initialize all descriptors */
446 		for (i = 0; i < ring->size; i++)
447 			mlx4_en_init_rx_desc(priv, ring, i);
448 
449 		/* Initialize page allocators */
450 		err = mlx4_en_init_allocator(priv, ring);
451 		if (err) {
452 			en_err(priv, "Failed initializing ring allocator\n");
453 			if (ring->stride <= TXBB_SIZE)
454 				ring->buf -= TXBB_SIZE;
455 			ring_ind--;
456 			goto err_allocator;
457 		}
458 	}
459 	err = mlx4_en_fill_rx_buffers(priv);
460 	if (err)
461 		goto err_buffers;
462 
463 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
464 		ring = priv->rx_ring[ring_ind];
465 
466 		ring->size_mask = ring->actual_size - 1;
467 		mlx4_en_update_rx_prod_db(ring);
468 	}
469 
470 	return 0;
471 
472 err_buffers:
473 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
474 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
475 
476 	ring_ind = priv->rx_ring_num - 1;
477 err_allocator:
478 	while (ring_ind >= 0) {
479 		if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
480 			priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
481 		mlx4_en_destroy_allocator(priv, priv->rx_ring[ring_ind]);
482 		ring_ind--;
483 	}
484 	return err;
485 }
486 
487 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
488 			     struct mlx4_en_rx_ring **pring,
489 			     u32 size, u16 stride)
490 {
491 	struct mlx4_en_dev *mdev = priv->mdev;
492 	struct mlx4_en_rx_ring *ring = *pring;
493 
494 	mlx4_en_unmap_buffer(&ring->wqres.buf);
495 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
496 	vfree(ring->rx_info);
497 	ring->rx_info = NULL;
498 	kfree(ring);
499 	*pring = NULL;
500 #ifdef CONFIG_RFS_ACCEL
501 	mlx4_en_cleanup_filters(priv);
502 #endif
503 }
504 
505 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
506 				struct mlx4_en_rx_ring *ring)
507 {
508 	mlx4_en_free_rx_buf(priv, ring);
509 	if (ring->stride <= TXBB_SIZE)
510 		ring->buf -= TXBB_SIZE;
511 	mlx4_en_destroy_allocator(priv, ring);
512 }
513 
514 
515 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
516 				    struct mlx4_en_rx_desc *rx_desc,
517 				    struct mlx4_en_rx_alloc *frags,
518 				    struct sk_buff *skb,
519 				    int length)
520 {
521 	struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
522 	struct mlx4_en_frag_info *frag_info;
523 	int nr;
524 	dma_addr_t dma;
525 
526 	/* Collect used fragments while replacing them in the HW descriptors */
527 	for (nr = 0; nr < priv->num_frags; nr++) {
528 		frag_info = &priv->frag_info[nr];
529 		if (length <= frag_info->frag_prefix_size)
530 			break;
531 		if (!frags[nr].page)
532 			goto fail;
533 
534 		dma = be64_to_cpu(rx_desc->data[nr].addr);
535 		dma_sync_single_for_cpu(priv->ddev, dma, frag_info->frag_size,
536 					DMA_FROM_DEVICE);
537 
538 		/* Save page reference in skb */
539 		__skb_frag_set_page(&skb_frags_rx[nr], frags[nr].page);
540 		skb_frag_size_set(&skb_frags_rx[nr], frag_info->frag_size);
541 		skb_frags_rx[nr].page_offset = frags[nr].page_offset;
542 		skb->truesize += frag_info->frag_stride;
543 		frags[nr].page = NULL;
544 	}
545 	/* Adjust size of last fragment to match actual length */
546 	if (nr > 0)
547 		skb_frag_size_set(&skb_frags_rx[nr - 1],
548 			length - priv->frag_info[nr - 1].frag_prefix_size);
549 	return nr;
550 
551 fail:
552 	while (nr > 0) {
553 		nr--;
554 		__skb_frag_unref(&skb_frags_rx[nr]);
555 	}
556 	return 0;
557 }
558 
559 
560 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
561 				      struct mlx4_en_rx_desc *rx_desc,
562 				      struct mlx4_en_rx_alloc *frags,
563 				      unsigned int length)
564 {
565 	struct sk_buff *skb;
566 	void *va;
567 	int used_frags;
568 	dma_addr_t dma;
569 
570 	skb = netdev_alloc_skb(priv->dev, SMALL_PACKET_SIZE + NET_IP_ALIGN);
571 	if (!skb) {
572 		en_dbg(RX_ERR, priv, "Failed allocating skb\n");
573 		return NULL;
574 	}
575 	skb_reserve(skb, NET_IP_ALIGN);
576 	skb->len = length;
577 
578 	/* Get pointer to first fragment so we could copy the headers into the
579 	 * (linear part of the) skb */
580 	va = page_address(frags[0].page) + frags[0].page_offset;
581 
582 	if (length <= SMALL_PACKET_SIZE) {
583 		/* We are copying all relevant data to the skb - temporarily
584 		 * sync buffers for the copy */
585 		dma = be64_to_cpu(rx_desc->data[0].addr);
586 		dma_sync_single_for_cpu(priv->ddev, dma, length,
587 					DMA_FROM_DEVICE);
588 		skb_copy_to_linear_data(skb, va, length);
589 		skb->tail += length;
590 	} else {
591 		/* Move relevant fragments to skb */
592 		used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, frags,
593 							skb, length);
594 		if (unlikely(!used_frags)) {
595 			kfree_skb(skb);
596 			return NULL;
597 		}
598 		skb_shinfo(skb)->nr_frags = used_frags;
599 
600 		/* Copy headers into the skb linear buffer */
601 		memcpy(skb->data, va, HEADER_COPY_SIZE);
602 		skb->tail += HEADER_COPY_SIZE;
603 
604 		/* Skip headers in first fragment */
605 		skb_shinfo(skb)->frags[0].page_offset += HEADER_COPY_SIZE;
606 
607 		/* Adjust size of first fragment */
608 		skb_frag_size_sub(&skb_shinfo(skb)->frags[0], HEADER_COPY_SIZE);
609 		skb->data_len = length - HEADER_COPY_SIZE;
610 	}
611 	return skb;
612 }
613 
614 static void validate_loopback(struct mlx4_en_priv *priv, struct sk_buff *skb)
615 {
616 	int i;
617 	int offset = ETH_HLEN;
618 
619 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
620 		if (*(skb->data + offset) != (unsigned char) (i & 0xff))
621 			goto out_loopback;
622 	}
623 	/* Loopback found */
624 	priv->loopback_ok = 1;
625 
626 out_loopback:
627 	dev_kfree_skb_any(skb);
628 }
629 
630 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
631 				     struct mlx4_en_rx_ring *ring)
632 {
633 	int index = ring->prod & ring->size_mask;
634 
635 	while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
636 		if (mlx4_en_prepare_rx_desc(priv, ring, index, GFP_ATOMIC))
637 			break;
638 		ring->prod++;
639 		index = ring->prod & ring->size_mask;
640 	}
641 }
642 
643 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
644 {
645 	struct mlx4_en_priv *priv = netdev_priv(dev);
646 	struct mlx4_en_dev *mdev = priv->mdev;
647 	struct mlx4_cqe *cqe;
648 	struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
649 	struct mlx4_en_rx_alloc *frags;
650 	struct mlx4_en_rx_desc *rx_desc;
651 	struct sk_buff *skb;
652 	int index;
653 	int nr;
654 	unsigned int length;
655 	int polled = 0;
656 	int ip_summed;
657 	int factor = priv->cqe_factor;
658 	u64 timestamp;
659 	bool l2_tunnel;
660 
661 	if (!priv->port_up)
662 		return 0;
663 
664 	if (budget <= 0)
665 		return polled;
666 
667 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
668 	 * descriptor offset can be deduced from the CQE index instead of
669 	 * reading 'cqe->index' */
670 	index = cq->mcq.cons_index & ring->size_mask;
671 	cqe = &cq->buf[(index << factor) + factor];
672 
673 	/* Process all completed CQEs */
674 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
675 		    cq->mcq.cons_index & cq->size)) {
676 
677 		frags = ring->rx_info + (index << priv->log_rx_info);
678 		rx_desc = ring->buf + (index << ring->log_stride);
679 
680 		/*
681 		 * make sure we read the CQE after we read the ownership bit
682 		 */
683 		rmb();
684 
685 		/* Drop packet on bad receive or bad checksum */
686 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
687 						MLX4_CQE_OPCODE_ERROR)) {
688 			en_err(priv, "CQE completed in error - vendor "
689 				  "syndrom:%d syndrom:%d\n",
690 				  ((struct mlx4_err_cqe *) cqe)->vendor_err_syndrome,
691 				  ((struct mlx4_err_cqe *) cqe)->syndrome);
692 			goto next;
693 		}
694 		if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
695 			en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
696 			goto next;
697 		}
698 
699 		/* Check if we need to drop the packet if SRIOV is not enabled
700 		 * and not performing the selftest or flb disabled
701 		 */
702 		if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
703 			struct ethhdr *ethh;
704 			dma_addr_t dma;
705 			/* Get pointer to first fragment since we haven't
706 			 * skb yet and cast it to ethhdr struct
707 			 */
708 			dma = be64_to_cpu(rx_desc->data[0].addr);
709 			dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
710 						DMA_FROM_DEVICE);
711 			ethh = (struct ethhdr *)(page_address(frags[0].page) +
712 						 frags[0].page_offset);
713 
714 			if (is_multicast_ether_addr(ethh->h_dest)) {
715 				struct mlx4_mac_entry *entry;
716 				struct hlist_head *bucket;
717 				unsigned int mac_hash;
718 
719 				/* Drop the packet, since HW loopback-ed it */
720 				mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
721 				bucket = &priv->mac_hash[mac_hash];
722 				rcu_read_lock();
723 				hlist_for_each_entry_rcu(entry, bucket, hlist) {
724 					if (ether_addr_equal_64bits(entry->mac,
725 								    ethh->h_source)) {
726 						rcu_read_unlock();
727 						goto next;
728 					}
729 				}
730 				rcu_read_unlock();
731 			}
732 		}
733 
734 		/*
735 		 * Packet is OK - process it.
736 		 */
737 		length = be32_to_cpu(cqe->byte_cnt);
738 		length -= ring->fcs_del;
739 		ring->bytes += length;
740 		ring->packets++;
741 		l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
742 			(cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
743 
744 		if (likely(dev->features & NETIF_F_RXCSUM)) {
745 			if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
746 			    (cqe->checksum == cpu_to_be16(0xffff))) {
747 				ring->csum_ok++;
748 				/* This packet is eligible for GRO if it is:
749 				 * - DIX Ethernet (type interpretation)
750 				 * - TCP/IP (v4)
751 				 * - without IP options
752 				 * - not an IP fragment
753 				 * - no LLS polling in progress
754 				 */
755 				if (!mlx4_en_cq_busy_polling(cq) &&
756 				    (dev->features & NETIF_F_GRO)) {
757 					struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
758 					if (!gro_skb)
759 						goto next;
760 
761 					nr = mlx4_en_complete_rx_desc(priv,
762 						rx_desc, frags, gro_skb,
763 						length);
764 					if (!nr)
765 						goto next;
766 
767 					skb_shinfo(gro_skb)->nr_frags = nr;
768 					gro_skb->len = length;
769 					gro_skb->data_len = length;
770 					gro_skb->ip_summed = CHECKSUM_UNNECESSARY;
771 
772 					if (l2_tunnel)
773 						gro_skb->encapsulation = 1;
774 					if ((cqe->vlan_my_qpn &
775 					    cpu_to_be32(MLX4_CQE_VLAN_PRESENT_MASK)) &&
776 					    (dev->features & NETIF_F_HW_VLAN_CTAG_RX)) {
777 						u16 vid = be16_to_cpu(cqe->sl_vid);
778 
779 						__vlan_hwaccel_put_tag(gro_skb, htons(ETH_P_8021Q), vid);
780 					}
781 
782 					if (dev->features & NETIF_F_RXHASH)
783 						skb_set_hash(gro_skb,
784 							     be32_to_cpu(cqe->immed_rss_invalid),
785 							     PKT_HASH_TYPE_L3);
786 
787 					skb_record_rx_queue(gro_skb, cq->ring);
788 
789 					if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
790 						timestamp = mlx4_en_get_cqe_ts(cqe);
791 						mlx4_en_fill_hwtstamps(mdev,
792 								       skb_hwtstamps(gro_skb),
793 								       timestamp);
794 					}
795 
796 					napi_gro_frags(&cq->napi);
797 					goto next;
798 				}
799 
800 				/* GRO not possible, complete processing here */
801 				ip_summed = CHECKSUM_UNNECESSARY;
802 			} else {
803 				ip_summed = CHECKSUM_NONE;
804 				ring->csum_none++;
805 			}
806 		} else {
807 			ip_summed = CHECKSUM_NONE;
808 			ring->csum_none++;
809 		}
810 
811 		skb = mlx4_en_rx_skb(priv, rx_desc, frags, length);
812 		if (!skb) {
813 			priv->stats.rx_dropped++;
814 			goto next;
815 		}
816 
817                 if (unlikely(priv->validate_loopback)) {
818 			validate_loopback(priv, skb);
819 			goto next;
820 		}
821 
822 		skb->ip_summed = ip_summed;
823 		skb->protocol = eth_type_trans(skb, dev);
824 		skb_record_rx_queue(skb, cq->ring);
825 
826 		if (l2_tunnel)
827 			skb->encapsulation = 1;
828 
829 		if (dev->features & NETIF_F_RXHASH)
830 			skb_set_hash(skb,
831 				     be32_to_cpu(cqe->immed_rss_invalid),
832 				     PKT_HASH_TYPE_L3);
833 
834 		if ((be32_to_cpu(cqe->vlan_my_qpn) &
835 		    MLX4_CQE_VLAN_PRESENT_MASK) &&
836 		    (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
837 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), be16_to_cpu(cqe->sl_vid));
838 
839 		if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
840 			timestamp = mlx4_en_get_cqe_ts(cqe);
841 			mlx4_en_fill_hwtstamps(mdev, skb_hwtstamps(skb),
842 					       timestamp);
843 		}
844 
845 		skb_mark_napi_id(skb, &cq->napi);
846 
847 		if (!mlx4_en_cq_busy_polling(cq))
848 			napi_gro_receive(&cq->napi, skb);
849 		else
850 			netif_receive_skb(skb);
851 
852 next:
853 		for (nr = 0; nr < priv->num_frags; nr++)
854 			mlx4_en_free_frag(priv, frags, nr);
855 
856 		++cq->mcq.cons_index;
857 		index = (cq->mcq.cons_index) & ring->size_mask;
858 		cqe = &cq->buf[(index << factor) + factor];
859 		if (++polled == budget)
860 			goto out;
861 	}
862 
863 out:
864 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
865 	mlx4_cq_set_ci(&cq->mcq);
866 	wmb(); /* ensure HW sees CQ consumer before we post new buffers */
867 	ring->cons = cq->mcq.cons_index;
868 	mlx4_en_refill_rx_buffers(priv, ring);
869 	mlx4_en_update_rx_prod_db(ring);
870 	return polled;
871 }
872 
873 
874 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
875 {
876 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
877 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
878 
879 	if (priv->port_up)
880 		napi_schedule(&cq->napi);
881 	else
882 		mlx4_en_arm_cq(priv, cq);
883 }
884 
885 /* Rx CQ polling - called by NAPI */
886 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
887 {
888 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
889 	struct net_device *dev = cq->dev;
890 	struct mlx4_en_priv *priv = netdev_priv(dev);
891 	int done;
892 
893 	if (!mlx4_en_cq_lock_napi(cq))
894 		return budget;
895 
896 	done = mlx4_en_process_rx_cq(dev, cq, budget);
897 
898 	mlx4_en_cq_unlock_napi(cq);
899 
900 	/* If we used up all the quota - we're probably not done yet... */
901 	if (done == budget)
902 		INC_PERF_COUNTER(priv->pstats.napi_quota);
903 	else {
904 		/* Done for now */
905 		napi_complete(napi);
906 		mlx4_en_arm_cq(priv, cq);
907 	}
908 	return done;
909 }
910 
911 static const int frag_sizes[] = {
912 	FRAG_SZ0,
913 	FRAG_SZ1,
914 	FRAG_SZ2,
915 	FRAG_SZ3
916 };
917 
918 void mlx4_en_calc_rx_buf(struct net_device *dev)
919 {
920 	struct mlx4_en_priv *priv = netdev_priv(dev);
921 	int eff_mtu = dev->mtu + ETH_HLEN + VLAN_HLEN + ETH_LLC_SNAP_SIZE;
922 	int buf_size = 0;
923 	int i = 0;
924 
925 	while (buf_size < eff_mtu) {
926 		priv->frag_info[i].frag_size =
927 			(eff_mtu > buf_size + frag_sizes[i]) ?
928 				frag_sizes[i] : eff_mtu - buf_size;
929 		priv->frag_info[i].frag_prefix_size = buf_size;
930 		if (!i)	{
931 			priv->frag_info[i].frag_align = NET_IP_ALIGN;
932 			priv->frag_info[i].frag_stride =
933 				ALIGN(frag_sizes[i] + NET_IP_ALIGN, SMP_CACHE_BYTES);
934 		} else {
935 			priv->frag_info[i].frag_align = 0;
936 			priv->frag_info[i].frag_stride =
937 				ALIGN(frag_sizes[i], SMP_CACHE_BYTES);
938 		}
939 		buf_size += priv->frag_info[i].frag_size;
940 		i++;
941 	}
942 
943 	priv->num_frags = i;
944 	priv->rx_skb_size = eff_mtu;
945 	priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
946 
947 	en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d "
948 		  "num_frags:%d):\n", eff_mtu, priv->num_frags);
949 	for (i = 0; i < priv->num_frags; i++) {
950 		en_err(priv,
951 		       "  frag:%d - size:%d prefix:%d align:%d stride:%d\n",
952 		       i,
953 		       priv->frag_info[i].frag_size,
954 		       priv->frag_info[i].frag_prefix_size,
955 		       priv->frag_info[i].frag_align,
956 		       priv->frag_info[i].frag_stride);
957 	}
958 }
959 
960 /* RSS related functions */
961 
962 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
963 				 struct mlx4_en_rx_ring *ring,
964 				 enum mlx4_qp_state *state,
965 				 struct mlx4_qp *qp)
966 {
967 	struct mlx4_en_dev *mdev = priv->mdev;
968 	struct mlx4_qp_context *context;
969 	int err = 0;
970 
971 	context = kmalloc(sizeof(*context), GFP_KERNEL);
972 	if (!context)
973 		return -ENOMEM;
974 
975 	err = mlx4_qp_alloc(mdev->dev, qpn, qp);
976 	if (err) {
977 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
978 		goto out;
979 	}
980 	qp->event = mlx4_en_sqp_event;
981 
982 	memset(context, 0, sizeof *context);
983 	mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
984 				qpn, ring->cqn, -1, context);
985 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
986 
987 	/* Cancel FCS removal if FW allows */
988 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
989 		context->param3 |= cpu_to_be32(1 << 29);
990 		ring->fcs_del = ETH_FCS_LEN;
991 	} else
992 		ring->fcs_del = 0;
993 
994 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
995 	if (err) {
996 		mlx4_qp_remove(mdev->dev, qp);
997 		mlx4_qp_free(mdev->dev, qp);
998 	}
999 	mlx4_en_update_rx_prod_db(ring);
1000 out:
1001 	kfree(context);
1002 	return err;
1003 }
1004 
1005 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
1006 {
1007 	int err;
1008 	u32 qpn;
1009 
1010 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn);
1011 	if (err) {
1012 		en_err(priv, "Failed reserving drop qpn\n");
1013 		return err;
1014 	}
1015 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp);
1016 	if (err) {
1017 		en_err(priv, "Failed allocating drop qp\n");
1018 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1019 		return err;
1020 	}
1021 
1022 	return 0;
1023 }
1024 
1025 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
1026 {
1027 	u32 qpn;
1028 
1029 	qpn = priv->drop_qp.qpn;
1030 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1031 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1032 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1033 }
1034 
1035 /* Allocate rx qp's and configure them according to rss map */
1036 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1037 {
1038 	struct mlx4_en_dev *mdev = priv->mdev;
1039 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1040 	struct mlx4_qp_context context;
1041 	struct mlx4_rss_context *rss_context;
1042 	int rss_rings;
1043 	void *ptr;
1044 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1045 			MLX4_RSS_TCP_IPV6);
1046 	int i, qpn;
1047 	int err = 0;
1048 	int good_qps = 0;
1049 	static const u32 rsskey[10] = { 0xD181C62C, 0xF7F4DB5B, 0x1983A2FC,
1050 				0x943E1ADB, 0xD9389E6B, 0xD1039C2C, 0xA74499AD,
1051 				0x593D56D9, 0xF3253C06, 0x2ADC1FFC};
1052 
1053 	en_dbg(DRV, priv, "Configuring rss steering\n");
1054 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1055 				    priv->rx_ring_num,
1056 				    &rss_map->base_qpn);
1057 	if (err) {
1058 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1059 		return err;
1060 	}
1061 
1062 	for (i = 0; i < priv->rx_ring_num; i++) {
1063 		qpn = rss_map->base_qpn + i;
1064 		err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1065 					    &rss_map->state[i],
1066 					    &rss_map->qps[i]);
1067 		if (err)
1068 			goto rss_err;
1069 
1070 		++good_qps;
1071 	}
1072 
1073 	/* Configure RSS indirection qp */
1074 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp);
1075 	if (err) {
1076 		en_err(priv, "Failed to allocate RSS indirection QP\n");
1077 		goto rss_err;
1078 	}
1079 	rss_map->indir_qp.event = mlx4_en_sqp_event;
1080 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1081 				priv->rx_ring[0]->cqn, -1, &context);
1082 
1083 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1084 		rss_rings = priv->rx_ring_num;
1085 	else
1086 		rss_rings = priv->prof->rss_rings;
1087 
1088 	ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1089 					+ MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1090 	rss_context = ptr;
1091 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1092 					    (rss_map->base_qpn));
1093 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1094 	if (priv->mdev->profile.udp_rss) {
1095 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1096 		rss_context->base_qpn_udp = rss_context->default_qpn;
1097 	}
1098 
1099 	if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1100 		en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1101 		rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1102 	}
1103 
1104 	rss_context->flags = rss_mask;
1105 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1106 	for (i = 0; i < 10; i++)
1107 		rss_context->rss_key[i] = cpu_to_be32(rsskey[i]);
1108 
1109 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1110 			       &rss_map->indir_qp, &rss_map->indir_state);
1111 	if (err)
1112 		goto indir_err;
1113 
1114 	return 0;
1115 
1116 indir_err:
1117 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1118 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1119 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1120 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1121 rss_err:
1122 	for (i = 0; i < good_qps; i++) {
1123 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1124 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1125 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1126 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1127 	}
1128 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1129 	return err;
1130 }
1131 
1132 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1133 {
1134 	struct mlx4_en_dev *mdev = priv->mdev;
1135 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1136 	int i;
1137 
1138 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1139 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1140 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1141 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1142 
1143 	for (i = 0; i < priv->rx_ring_num; i++) {
1144 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1145 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1146 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1147 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1148 	}
1149 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1150 }
1151