xref: /linux/drivers/net/ethernet/mellanox/mlx4/en_rx.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 #include <linux/irq.h>
44 
45 #if IS_ENABLED(CONFIG_IPV6)
46 #include <net/ip6_checksum.h>
47 #endif
48 
49 #include "mlx4_en.h"
50 
51 static int mlx4_alloc_pages(struct mlx4_en_priv *priv,
52 			    struct mlx4_en_rx_alloc *page_alloc,
53 			    const struct mlx4_en_frag_info *frag_info,
54 			    gfp_t _gfp)
55 {
56 	int order;
57 	struct page *page;
58 	dma_addr_t dma;
59 
60 	for (order = MLX4_EN_ALLOC_PREFER_ORDER; ;) {
61 		gfp_t gfp = _gfp;
62 
63 		if (order)
64 			gfp |= __GFP_COMP | __GFP_NOWARN;
65 		page = alloc_pages(gfp, order);
66 		if (likely(page))
67 			break;
68 		if (--order < 0 ||
69 		    ((PAGE_SIZE << order) < frag_info->frag_size))
70 			return -ENOMEM;
71 	}
72 	dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE << order,
73 			   PCI_DMA_FROMDEVICE);
74 	if (dma_mapping_error(priv->ddev, dma)) {
75 		put_page(page);
76 		return -ENOMEM;
77 	}
78 	page_alloc->page_size = PAGE_SIZE << order;
79 	page_alloc->page = page;
80 	page_alloc->dma = dma;
81 	page_alloc->page_offset = 0;
82 	/* Not doing get_page() for each frag is a big win
83 	 * on asymetric workloads. Note we can not use atomic_set().
84 	 */
85 	atomic_add(page_alloc->page_size / frag_info->frag_stride - 1,
86 		   &page->_count);
87 	return 0;
88 }
89 
90 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
91 			       struct mlx4_en_rx_desc *rx_desc,
92 			       struct mlx4_en_rx_alloc *frags,
93 			       struct mlx4_en_rx_alloc *ring_alloc,
94 			       gfp_t gfp)
95 {
96 	struct mlx4_en_rx_alloc page_alloc[MLX4_EN_MAX_RX_FRAGS];
97 	const struct mlx4_en_frag_info *frag_info;
98 	struct page *page;
99 	dma_addr_t dma;
100 	int i;
101 
102 	for (i = 0; i < priv->num_frags; i++) {
103 		frag_info = &priv->frag_info[i];
104 		page_alloc[i] = ring_alloc[i];
105 		page_alloc[i].page_offset += frag_info->frag_stride;
106 
107 		if (page_alloc[i].page_offset + frag_info->frag_stride <=
108 		    ring_alloc[i].page_size)
109 			continue;
110 
111 		if (mlx4_alloc_pages(priv, &page_alloc[i], frag_info, gfp))
112 			goto out;
113 	}
114 
115 	for (i = 0; i < priv->num_frags; i++) {
116 		frags[i] = ring_alloc[i];
117 		dma = ring_alloc[i].dma + ring_alloc[i].page_offset;
118 		ring_alloc[i] = page_alloc[i];
119 		rx_desc->data[i].addr = cpu_to_be64(dma);
120 	}
121 
122 	return 0;
123 
124 out:
125 	while (i--) {
126 		if (page_alloc[i].page != ring_alloc[i].page) {
127 			dma_unmap_page(priv->ddev, page_alloc[i].dma,
128 				page_alloc[i].page_size, PCI_DMA_FROMDEVICE);
129 			page = page_alloc[i].page;
130 			atomic_set(&page->_count, 1);
131 			put_page(page);
132 		}
133 	}
134 	return -ENOMEM;
135 }
136 
137 static void mlx4_en_free_frag(struct mlx4_en_priv *priv,
138 			      struct mlx4_en_rx_alloc *frags,
139 			      int i)
140 {
141 	const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
142 	u32 next_frag_end = frags[i].page_offset + 2 * frag_info->frag_stride;
143 
144 
145 	if (next_frag_end > frags[i].page_size)
146 		dma_unmap_page(priv->ddev, frags[i].dma, frags[i].page_size,
147 			       PCI_DMA_FROMDEVICE);
148 
149 	if (frags[i].page)
150 		put_page(frags[i].page);
151 }
152 
153 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
154 				  struct mlx4_en_rx_ring *ring)
155 {
156 	int i;
157 	struct mlx4_en_rx_alloc *page_alloc;
158 
159 	for (i = 0; i < priv->num_frags; i++) {
160 		const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
161 
162 		if (mlx4_alloc_pages(priv, &ring->page_alloc[i],
163 				     frag_info, GFP_KERNEL | __GFP_COLD))
164 			goto out;
165 
166 		en_dbg(DRV, priv, "  frag %d allocator: - size:%d frags:%d\n",
167 		       i, ring->page_alloc[i].page_size,
168 		       atomic_read(&ring->page_alloc[i].page->_count));
169 	}
170 	return 0;
171 
172 out:
173 	while (i--) {
174 		struct page *page;
175 
176 		page_alloc = &ring->page_alloc[i];
177 		dma_unmap_page(priv->ddev, page_alloc->dma,
178 			       page_alloc->page_size, PCI_DMA_FROMDEVICE);
179 		page = page_alloc->page;
180 		atomic_set(&page->_count, 1);
181 		put_page(page);
182 		page_alloc->page = NULL;
183 	}
184 	return -ENOMEM;
185 }
186 
187 static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
188 				      struct mlx4_en_rx_ring *ring)
189 {
190 	struct mlx4_en_rx_alloc *page_alloc;
191 	int i;
192 
193 	for (i = 0; i < priv->num_frags; i++) {
194 		const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
195 
196 		page_alloc = &ring->page_alloc[i];
197 		en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
198 		       i, page_count(page_alloc->page));
199 
200 		dma_unmap_page(priv->ddev, page_alloc->dma,
201 				page_alloc->page_size, PCI_DMA_FROMDEVICE);
202 		while (page_alloc->page_offset + frag_info->frag_stride <
203 		       page_alloc->page_size) {
204 			put_page(page_alloc->page);
205 			page_alloc->page_offset += frag_info->frag_stride;
206 		}
207 		page_alloc->page = NULL;
208 	}
209 }
210 
211 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
212 				 struct mlx4_en_rx_ring *ring, int index)
213 {
214 	struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
215 	int possible_frags;
216 	int i;
217 
218 	/* Set size and memtype fields */
219 	for (i = 0; i < priv->num_frags; i++) {
220 		rx_desc->data[i].byte_count =
221 			cpu_to_be32(priv->frag_info[i].frag_size);
222 		rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
223 	}
224 
225 	/* If the number of used fragments does not fill up the ring stride,
226 	 * remaining (unused) fragments must be padded with null address/size
227 	 * and a special memory key */
228 	possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
229 	for (i = priv->num_frags; i < possible_frags; i++) {
230 		rx_desc->data[i].byte_count = 0;
231 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
232 		rx_desc->data[i].addr = 0;
233 	}
234 }
235 
236 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
237 				   struct mlx4_en_rx_ring *ring, int index,
238 				   gfp_t gfp)
239 {
240 	struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
241 	struct mlx4_en_rx_alloc *frags = ring->rx_info +
242 					(index << priv->log_rx_info);
243 
244 	return mlx4_en_alloc_frags(priv, rx_desc, frags, ring->page_alloc, gfp);
245 }
246 
247 static inline bool mlx4_en_is_ring_empty(struct mlx4_en_rx_ring *ring)
248 {
249 	return ring->prod == ring->cons;
250 }
251 
252 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
253 {
254 	*ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
255 }
256 
257 static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
258 				 struct mlx4_en_rx_ring *ring,
259 				 int index)
260 {
261 	struct mlx4_en_rx_alloc *frags;
262 	int nr;
263 
264 	frags = ring->rx_info + (index << priv->log_rx_info);
265 	for (nr = 0; nr < priv->num_frags; nr++) {
266 		en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
267 		mlx4_en_free_frag(priv, frags, nr);
268 	}
269 }
270 
271 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
272 {
273 	struct mlx4_en_rx_ring *ring;
274 	int ring_ind;
275 	int buf_ind;
276 	int new_size;
277 
278 	for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
279 		for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
280 			ring = priv->rx_ring[ring_ind];
281 
282 			if (mlx4_en_prepare_rx_desc(priv, ring,
283 						    ring->actual_size,
284 						    GFP_KERNEL | __GFP_COLD)) {
285 				if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
286 					en_err(priv, "Failed to allocate enough rx buffers\n");
287 					return -ENOMEM;
288 				} else {
289 					new_size = rounddown_pow_of_two(ring->actual_size);
290 					en_warn(priv, "Only %d buffers allocated reducing ring size to %d\n",
291 						ring->actual_size, new_size);
292 					goto reduce_rings;
293 				}
294 			}
295 			ring->actual_size++;
296 			ring->prod++;
297 		}
298 	}
299 	return 0;
300 
301 reduce_rings:
302 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
303 		ring = priv->rx_ring[ring_ind];
304 		while (ring->actual_size > new_size) {
305 			ring->actual_size--;
306 			ring->prod--;
307 			mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
308 		}
309 	}
310 
311 	return 0;
312 }
313 
314 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
315 				struct mlx4_en_rx_ring *ring)
316 {
317 	int index;
318 
319 	en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
320 	       ring->cons, ring->prod);
321 
322 	/* Unmap and free Rx buffers */
323 	while (!mlx4_en_is_ring_empty(ring)) {
324 		index = ring->cons & ring->size_mask;
325 		en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
326 		mlx4_en_free_rx_desc(priv, ring, index);
327 		++ring->cons;
328 	}
329 }
330 
331 void mlx4_en_set_num_rx_rings(struct mlx4_en_dev *mdev)
332 {
333 	int i;
334 	int num_of_eqs;
335 	int num_rx_rings;
336 	struct mlx4_dev *dev = mdev->dev;
337 
338 	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH) {
339 		num_of_eqs = max_t(int, MIN_RX_RINGS,
340 				   min_t(int,
341 					 mlx4_get_eqs_per_port(mdev->dev, i),
342 					 DEF_RX_RINGS));
343 
344 		num_rx_rings = mlx4_low_memory_profile() ? MIN_RX_RINGS :
345 			min_t(int, num_of_eqs,
346 			      netif_get_num_default_rss_queues());
347 		mdev->profile.prof[i].rx_ring_num =
348 			rounddown_pow_of_two(num_rx_rings);
349 	}
350 }
351 
352 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
353 			   struct mlx4_en_rx_ring **pring,
354 			   u32 size, u16 stride, int node)
355 {
356 	struct mlx4_en_dev *mdev = priv->mdev;
357 	struct mlx4_en_rx_ring *ring;
358 	int err = -ENOMEM;
359 	int tmp;
360 
361 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
362 	if (!ring) {
363 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
364 		if (!ring) {
365 			en_err(priv, "Failed to allocate RX ring structure\n");
366 			return -ENOMEM;
367 		}
368 	}
369 
370 	ring->prod = 0;
371 	ring->cons = 0;
372 	ring->size = size;
373 	ring->size_mask = size - 1;
374 	ring->stride = stride;
375 	ring->log_stride = ffs(ring->stride) - 1;
376 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
377 
378 	tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
379 					sizeof(struct mlx4_en_rx_alloc));
380 	ring->rx_info = vmalloc_node(tmp, node);
381 	if (!ring->rx_info) {
382 		ring->rx_info = vmalloc(tmp);
383 		if (!ring->rx_info) {
384 			err = -ENOMEM;
385 			goto err_ring;
386 		}
387 	}
388 
389 	en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
390 		 ring->rx_info, tmp);
391 
392 	/* Allocate HW buffers on provided NUMA node */
393 	set_dev_node(&mdev->dev->persist->pdev->dev, node);
394 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres,
395 				 ring->buf_size, 2 * PAGE_SIZE);
396 	set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
397 	if (err)
398 		goto err_info;
399 
400 	err = mlx4_en_map_buffer(&ring->wqres.buf);
401 	if (err) {
402 		en_err(priv, "Failed to map RX buffer\n");
403 		goto err_hwq;
404 	}
405 	ring->buf = ring->wqres.buf.direct.buf;
406 
407 	ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
408 
409 	*pring = ring;
410 	return 0;
411 
412 err_hwq:
413 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
414 err_info:
415 	vfree(ring->rx_info);
416 	ring->rx_info = NULL;
417 err_ring:
418 	kfree(ring);
419 	*pring = NULL;
420 
421 	return err;
422 }
423 
424 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
425 {
426 	struct mlx4_en_rx_ring *ring;
427 	int i;
428 	int ring_ind;
429 	int err;
430 	int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
431 					DS_SIZE * priv->num_frags);
432 
433 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
434 		ring = priv->rx_ring[ring_ind];
435 
436 		ring->prod = 0;
437 		ring->cons = 0;
438 		ring->actual_size = 0;
439 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
440 
441 		ring->stride = stride;
442 		if (ring->stride <= TXBB_SIZE)
443 			ring->buf += TXBB_SIZE;
444 
445 		ring->log_stride = ffs(ring->stride) - 1;
446 		ring->buf_size = ring->size * ring->stride;
447 
448 		memset(ring->buf, 0, ring->buf_size);
449 		mlx4_en_update_rx_prod_db(ring);
450 
451 		/* Initialize all descriptors */
452 		for (i = 0; i < ring->size; i++)
453 			mlx4_en_init_rx_desc(priv, ring, i);
454 
455 		/* Initialize page allocators */
456 		err = mlx4_en_init_allocator(priv, ring);
457 		if (err) {
458 			en_err(priv, "Failed initializing ring allocator\n");
459 			if (ring->stride <= TXBB_SIZE)
460 				ring->buf -= TXBB_SIZE;
461 			ring_ind--;
462 			goto err_allocator;
463 		}
464 	}
465 	err = mlx4_en_fill_rx_buffers(priv);
466 	if (err)
467 		goto err_buffers;
468 
469 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
470 		ring = priv->rx_ring[ring_ind];
471 
472 		ring->size_mask = ring->actual_size - 1;
473 		mlx4_en_update_rx_prod_db(ring);
474 	}
475 
476 	return 0;
477 
478 err_buffers:
479 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
480 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
481 
482 	ring_ind = priv->rx_ring_num - 1;
483 err_allocator:
484 	while (ring_ind >= 0) {
485 		if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
486 			priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
487 		mlx4_en_destroy_allocator(priv, priv->rx_ring[ring_ind]);
488 		ring_ind--;
489 	}
490 	return err;
491 }
492 
493 /* We recover from out of memory by scheduling our napi poll
494  * function (mlx4_en_process_cq), which tries to allocate
495  * all missing RX buffers (call to mlx4_en_refill_rx_buffers).
496  */
497 void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv)
498 {
499 	int ring;
500 
501 	if (!priv->port_up)
502 		return;
503 
504 	for (ring = 0; ring < priv->rx_ring_num; ring++) {
505 		if (mlx4_en_is_ring_empty(priv->rx_ring[ring]))
506 			napi_reschedule(&priv->rx_cq[ring]->napi);
507 	}
508 }
509 
510 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
511 			     struct mlx4_en_rx_ring **pring,
512 			     u32 size, u16 stride)
513 {
514 	struct mlx4_en_dev *mdev = priv->mdev;
515 	struct mlx4_en_rx_ring *ring = *pring;
516 
517 	mlx4_en_unmap_buffer(&ring->wqres.buf);
518 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
519 	vfree(ring->rx_info);
520 	ring->rx_info = NULL;
521 	kfree(ring);
522 	*pring = NULL;
523 #ifdef CONFIG_RFS_ACCEL
524 	mlx4_en_cleanup_filters(priv);
525 #endif
526 }
527 
528 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
529 				struct mlx4_en_rx_ring *ring)
530 {
531 	mlx4_en_free_rx_buf(priv, ring);
532 	if (ring->stride <= TXBB_SIZE)
533 		ring->buf -= TXBB_SIZE;
534 	mlx4_en_destroy_allocator(priv, ring);
535 }
536 
537 
538 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
539 				    struct mlx4_en_rx_desc *rx_desc,
540 				    struct mlx4_en_rx_alloc *frags,
541 				    struct sk_buff *skb,
542 				    int length)
543 {
544 	struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
545 	struct mlx4_en_frag_info *frag_info;
546 	int nr;
547 	dma_addr_t dma;
548 
549 	/* Collect used fragments while replacing them in the HW descriptors */
550 	for (nr = 0; nr < priv->num_frags; nr++) {
551 		frag_info = &priv->frag_info[nr];
552 		if (length <= frag_info->frag_prefix_size)
553 			break;
554 		if (!frags[nr].page)
555 			goto fail;
556 
557 		dma = be64_to_cpu(rx_desc->data[nr].addr);
558 		dma_sync_single_for_cpu(priv->ddev, dma, frag_info->frag_size,
559 					DMA_FROM_DEVICE);
560 
561 		/* Save page reference in skb */
562 		__skb_frag_set_page(&skb_frags_rx[nr], frags[nr].page);
563 		skb_frag_size_set(&skb_frags_rx[nr], frag_info->frag_size);
564 		skb_frags_rx[nr].page_offset = frags[nr].page_offset;
565 		skb->truesize += frag_info->frag_stride;
566 		frags[nr].page = NULL;
567 	}
568 	/* Adjust size of last fragment to match actual length */
569 	if (nr > 0)
570 		skb_frag_size_set(&skb_frags_rx[nr - 1],
571 			length - priv->frag_info[nr - 1].frag_prefix_size);
572 	return nr;
573 
574 fail:
575 	while (nr > 0) {
576 		nr--;
577 		__skb_frag_unref(&skb_frags_rx[nr]);
578 	}
579 	return 0;
580 }
581 
582 
583 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
584 				      struct mlx4_en_rx_desc *rx_desc,
585 				      struct mlx4_en_rx_alloc *frags,
586 				      unsigned int length)
587 {
588 	struct sk_buff *skb;
589 	void *va;
590 	int used_frags;
591 	dma_addr_t dma;
592 
593 	skb = netdev_alloc_skb(priv->dev, SMALL_PACKET_SIZE + NET_IP_ALIGN);
594 	if (!skb) {
595 		en_dbg(RX_ERR, priv, "Failed allocating skb\n");
596 		return NULL;
597 	}
598 	skb_reserve(skb, NET_IP_ALIGN);
599 	skb->len = length;
600 
601 	/* Get pointer to first fragment so we could copy the headers into the
602 	 * (linear part of the) skb */
603 	va = page_address(frags[0].page) + frags[0].page_offset;
604 
605 	if (length <= SMALL_PACKET_SIZE) {
606 		/* We are copying all relevant data to the skb - temporarily
607 		 * sync buffers for the copy */
608 		dma = be64_to_cpu(rx_desc->data[0].addr);
609 		dma_sync_single_for_cpu(priv->ddev, dma, length,
610 					DMA_FROM_DEVICE);
611 		skb_copy_to_linear_data(skb, va, length);
612 		skb->tail += length;
613 	} else {
614 		unsigned int pull_len;
615 
616 		/* Move relevant fragments to skb */
617 		used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, frags,
618 							skb, length);
619 		if (unlikely(!used_frags)) {
620 			kfree_skb(skb);
621 			return NULL;
622 		}
623 		skb_shinfo(skb)->nr_frags = used_frags;
624 
625 		pull_len = eth_get_headlen(va, SMALL_PACKET_SIZE);
626 		/* Copy headers into the skb linear buffer */
627 		memcpy(skb->data, va, pull_len);
628 		skb->tail += pull_len;
629 
630 		/* Skip headers in first fragment */
631 		skb_shinfo(skb)->frags[0].page_offset += pull_len;
632 
633 		/* Adjust size of first fragment */
634 		skb_frag_size_sub(&skb_shinfo(skb)->frags[0], pull_len);
635 		skb->data_len = length - pull_len;
636 	}
637 	return skb;
638 }
639 
640 static void validate_loopback(struct mlx4_en_priv *priv, struct sk_buff *skb)
641 {
642 	int i;
643 	int offset = ETH_HLEN;
644 
645 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
646 		if (*(skb->data + offset) != (unsigned char) (i & 0xff))
647 			goto out_loopback;
648 	}
649 	/* Loopback found */
650 	priv->loopback_ok = 1;
651 
652 out_loopback:
653 	dev_kfree_skb_any(skb);
654 }
655 
656 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
657 				     struct mlx4_en_rx_ring *ring)
658 {
659 	int index = ring->prod & ring->size_mask;
660 
661 	while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
662 		if (mlx4_en_prepare_rx_desc(priv, ring, index,
663 					    GFP_ATOMIC | __GFP_COLD))
664 			break;
665 		ring->prod++;
666 		index = ring->prod & ring->size_mask;
667 	}
668 }
669 
670 /* When hardware doesn't strip the vlan, we need to calculate the checksum
671  * over it and add it to the hardware's checksum calculation
672  */
673 static inline __wsum get_fixed_vlan_csum(__wsum hw_checksum,
674 					 struct vlan_hdr *vlanh)
675 {
676 	return csum_add(hw_checksum, *(__wsum *)vlanh);
677 }
678 
679 /* Although the stack expects checksum which doesn't include the pseudo
680  * header, the HW adds it. To address that, we are subtracting the pseudo
681  * header checksum from the checksum value provided by the HW.
682  */
683 static void get_fixed_ipv4_csum(__wsum hw_checksum, struct sk_buff *skb,
684 				struct iphdr *iph)
685 {
686 	__u16 length_for_csum = 0;
687 	__wsum csum_pseudo_header = 0;
688 
689 	length_for_csum = (be16_to_cpu(iph->tot_len) - (iph->ihl << 2));
690 	csum_pseudo_header = csum_tcpudp_nofold(iph->saddr, iph->daddr,
691 						length_for_csum, iph->protocol, 0);
692 	skb->csum = csum_sub(hw_checksum, csum_pseudo_header);
693 }
694 
695 #if IS_ENABLED(CONFIG_IPV6)
696 /* In IPv6 packets, besides subtracting the pseudo header checksum,
697  * we also compute/add the IP header checksum which
698  * is not added by the HW.
699  */
700 static int get_fixed_ipv6_csum(__wsum hw_checksum, struct sk_buff *skb,
701 			       struct ipv6hdr *ipv6h)
702 {
703 	__wsum csum_pseudo_hdr = 0;
704 
705 	if (ipv6h->nexthdr == IPPROTO_FRAGMENT || ipv6h->nexthdr == IPPROTO_HOPOPTS)
706 		return -1;
707 	hw_checksum = csum_add(hw_checksum, (__force __wsum)(ipv6h->nexthdr << 8));
708 
709 	csum_pseudo_hdr = csum_partial(&ipv6h->saddr,
710 				       sizeof(ipv6h->saddr) + sizeof(ipv6h->daddr), 0);
711 	csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ipv6h->payload_len);
712 	csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ntohs(ipv6h->nexthdr));
713 
714 	skb->csum = csum_sub(hw_checksum, csum_pseudo_hdr);
715 	skb->csum = csum_add(skb->csum, csum_partial(ipv6h, sizeof(struct ipv6hdr), 0));
716 	return 0;
717 }
718 #endif
719 static int check_csum(struct mlx4_cqe *cqe, struct sk_buff *skb, void *va,
720 		      netdev_features_t dev_features)
721 {
722 	__wsum hw_checksum = 0;
723 
724 	void *hdr = (u8 *)va + sizeof(struct ethhdr);
725 
726 	hw_checksum = csum_unfold((__force __sum16)cqe->checksum);
727 
728 	if (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK) &&
729 	    !(dev_features & NETIF_F_HW_VLAN_CTAG_RX)) {
730 		hw_checksum = get_fixed_vlan_csum(hw_checksum, hdr);
731 		hdr += sizeof(struct vlan_hdr);
732 	}
733 
734 	if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4))
735 		get_fixed_ipv4_csum(hw_checksum, skb, hdr);
736 #if IS_ENABLED(CONFIG_IPV6)
737 	else if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV6))
738 		if (get_fixed_ipv6_csum(hw_checksum, skb, hdr))
739 			return -1;
740 #endif
741 	return 0;
742 }
743 
744 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
745 {
746 	struct mlx4_en_priv *priv = netdev_priv(dev);
747 	struct mlx4_en_dev *mdev = priv->mdev;
748 	struct mlx4_cqe *cqe;
749 	struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
750 	struct mlx4_en_rx_alloc *frags;
751 	struct mlx4_en_rx_desc *rx_desc;
752 	struct sk_buff *skb;
753 	int index;
754 	int nr;
755 	unsigned int length;
756 	int polled = 0;
757 	int ip_summed;
758 	int factor = priv->cqe_factor;
759 	u64 timestamp;
760 	bool l2_tunnel;
761 
762 	if (!priv->port_up)
763 		return 0;
764 
765 	if (budget <= 0)
766 		return polled;
767 
768 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
769 	 * descriptor offset can be deduced from the CQE index instead of
770 	 * reading 'cqe->index' */
771 	index = cq->mcq.cons_index & ring->size_mask;
772 	cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
773 
774 	/* Process all completed CQEs */
775 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
776 		    cq->mcq.cons_index & cq->size)) {
777 
778 		frags = ring->rx_info + (index << priv->log_rx_info);
779 		rx_desc = ring->buf + (index << ring->log_stride);
780 
781 		/*
782 		 * make sure we read the CQE after we read the ownership bit
783 		 */
784 		dma_rmb();
785 
786 		/* Drop packet on bad receive or bad checksum */
787 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
788 						MLX4_CQE_OPCODE_ERROR)) {
789 			en_err(priv, "CQE completed in error - vendor syndrom:%d syndrom:%d\n",
790 			       ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
791 			       ((struct mlx4_err_cqe *)cqe)->syndrome);
792 			goto next;
793 		}
794 		if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
795 			en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
796 			goto next;
797 		}
798 
799 		/* Check if we need to drop the packet if SRIOV is not enabled
800 		 * and not performing the selftest or flb disabled
801 		 */
802 		if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
803 			struct ethhdr *ethh;
804 			dma_addr_t dma;
805 			/* Get pointer to first fragment since we haven't
806 			 * skb yet and cast it to ethhdr struct
807 			 */
808 			dma = be64_to_cpu(rx_desc->data[0].addr);
809 			dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
810 						DMA_FROM_DEVICE);
811 			ethh = (struct ethhdr *)(page_address(frags[0].page) +
812 						 frags[0].page_offset);
813 
814 			if (is_multicast_ether_addr(ethh->h_dest)) {
815 				struct mlx4_mac_entry *entry;
816 				struct hlist_head *bucket;
817 				unsigned int mac_hash;
818 
819 				/* Drop the packet, since HW loopback-ed it */
820 				mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
821 				bucket = &priv->mac_hash[mac_hash];
822 				rcu_read_lock();
823 				hlist_for_each_entry_rcu(entry, bucket, hlist) {
824 					if (ether_addr_equal_64bits(entry->mac,
825 								    ethh->h_source)) {
826 						rcu_read_unlock();
827 						goto next;
828 					}
829 				}
830 				rcu_read_unlock();
831 			}
832 		}
833 
834 		/*
835 		 * Packet is OK - process it.
836 		 */
837 		length = be32_to_cpu(cqe->byte_cnt);
838 		length -= ring->fcs_del;
839 		ring->bytes += length;
840 		ring->packets++;
841 		l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
842 			(cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
843 
844 		if (likely(dev->features & NETIF_F_RXCSUM)) {
845 			if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
846 						      MLX4_CQE_STATUS_UDP)) {
847 				if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
848 				    cqe->checksum == cpu_to_be16(0xffff)) {
849 					ip_summed = CHECKSUM_UNNECESSARY;
850 					ring->csum_ok++;
851 				} else {
852 					ip_summed = CHECKSUM_NONE;
853 					ring->csum_none++;
854 				}
855 			} else {
856 				if (priv->flags & MLX4_EN_FLAG_RX_CSUM_NON_TCP_UDP &&
857 				    (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
858 							       MLX4_CQE_STATUS_IPV6))) {
859 					ip_summed = CHECKSUM_COMPLETE;
860 					ring->csum_complete++;
861 				} else {
862 					ip_summed = CHECKSUM_NONE;
863 					ring->csum_none++;
864 				}
865 			}
866 		} else {
867 			ip_summed = CHECKSUM_NONE;
868 			ring->csum_none++;
869 		}
870 
871 		/* This packet is eligible for GRO if it is:
872 		 * - DIX Ethernet (type interpretation)
873 		 * - TCP/IP (v4)
874 		 * - without IP options
875 		 * - not an IP fragment
876 		 * - no LLS polling in progress
877 		 */
878 		if (!mlx4_en_cq_busy_polling(cq) &&
879 		    (dev->features & NETIF_F_GRO)) {
880 			struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
881 			if (!gro_skb)
882 				goto next;
883 
884 			nr = mlx4_en_complete_rx_desc(priv,
885 				rx_desc, frags, gro_skb,
886 				length);
887 			if (!nr)
888 				goto next;
889 
890 			if (ip_summed == CHECKSUM_COMPLETE) {
891 				void *va = skb_frag_address(skb_shinfo(gro_skb)->frags);
892 				if (check_csum(cqe, gro_skb, va,
893 					       dev->features)) {
894 					ip_summed = CHECKSUM_NONE;
895 					ring->csum_none++;
896 					ring->csum_complete--;
897 				}
898 			}
899 
900 			skb_shinfo(gro_skb)->nr_frags = nr;
901 			gro_skb->len = length;
902 			gro_skb->data_len = length;
903 			gro_skb->ip_summed = ip_summed;
904 
905 			if (l2_tunnel && ip_summed == CHECKSUM_UNNECESSARY)
906 				gro_skb->csum_level = 1;
907 
908 			if ((cqe->vlan_my_qpn &
909 			    cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK)) &&
910 			    (dev->features & NETIF_F_HW_VLAN_CTAG_RX)) {
911 				u16 vid = be16_to_cpu(cqe->sl_vid);
912 
913 				__vlan_hwaccel_put_tag(gro_skb, htons(ETH_P_8021Q), vid);
914 			} else if ((be32_to_cpu(cqe->vlan_my_qpn) &
915 				  MLX4_CQE_SVLAN_PRESENT_MASK) &&
916 				 (dev->features & NETIF_F_HW_VLAN_STAG_RX)) {
917 				__vlan_hwaccel_put_tag(gro_skb,
918 						       htons(ETH_P_8021AD),
919 						       be16_to_cpu(cqe->sl_vid));
920 			}
921 
922 			if (dev->features & NETIF_F_RXHASH)
923 				skb_set_hash(gro_skb,
924 					     be32_to_cpu(cqe->immed_rss_invalid),
925 					     (ip_summed == CHECKSUM_UNNECESSARY) ?
926 						PKT_HASH_TYPE_L4 :
927 						PKT_HASH_TYPE_L3);
928 
929 			skb_record_rx_queue(gro_skb, cq->ring);
930 			skb_mark_napi_id(gro_skb, &cq->napi);
931 
932 			if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
933 				timestamp = mlx4_en_get_cqe_ts(cqe);
934 				mlx4_en_fill_hwtstamps(mdev,
935 						       skb_hwtstamps(gro_skb),
936 						       timestamp);
937 			}
938 
939 			napi_gro_frags(&cq->napi);
940 			goto next;
941 		}
942 
943 		/* GRO not possible, complete processing here */
944 		skb = mlx4_en_rx_skb(priv, rx_desc, frags, length);
945 		if (!skb) {
946 			priv->stats.rx_dropped++;
947 			goto next;
948 		}
949 
950                 if (unlikely(priv->validate_loopback)) {
951 			validate_loopback(priv, skb);
952 			goto next;
953 		}
954 
955 		if (ip_summed == CHECKSUM_COMPLETE) {
956 			if (check_csum(cqe, skb, skb->data, dev->features)) {
957 				ip_summed = CHECKSUM_NONE;
958 				ring->csum_complete--;
959 				ring->csum_none++;
960 			}
961 		}
962 
963 		skb->ip_summed = ip_summed;
964 		skb->protocol = eth_type_trans(skb, dev);
965 		skb_record_rx_queue(skb, cq->ring);
966 
967 		if (l2_tunnel && ip_summed == CHECKSUM_UNNECESSARY)
968 			skb->csum_level = 1;
969 
970 		if (dev->features & NETIF_F_RXHASH)
971 			skb_set_hash(skb,
972 				     be32_to_cpu(cqe->immed_rss_invalid),
973 				     (ip_summed == CHECKSUM_UNNECESSARY) ?
974 					PKT_HASH_TYPE_L4 :
975 					PKT_HASH_TYPE_L3);
976 
977 		if ((be32_to_cpu(cqe->vlan_my_qpn) &
978 		    MLX4_CQE_CVLAN_PRESENT_MASK) &&
979 		    (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
980 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), be16_to_cpu(cqe->sl_vid));
981 		else if ((be32_to_cpu(cqe->vlan_my_qpn) &
982 			  MLX4_CQE_SVLAN_PRESENT_MASK) &&
983 			 (dev->features & NETIF_F_HW_VLAN_STAG_RX))
984 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021AD),
985 					       be16_to_cpu(cqe->sl_vid));
986 
987 		if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
988 			timestamp = mlx4_en_get_cqe_ts(cqe);
989 			mlx4_en_fill_hwtstamps(mdev, skb_hwtstamps(skb),
990 					       timestamp);
991 		}
992 
993 		skb_mark_napi_id(skb, &cq->napi);
994 
995 		if (!mlx4_en_cq_busy_polling(cq))
996 			napi_gro_receive(&cq->napi, skb);
997 		else
998 			netif_receive_skb(skb);
999 
1000 next:
1001 		for (nr = 0; nr < priv->num_frags; nr++)
1002 			mlx4_en_free_frag(priv, frags, nr);
1003 
1004 		++cq->mcq.cons_index;
1005 		index = (cq->mcq.cons_index) & ring->size_mask;
1006 		cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
1007 		if (++polled == budget)
1008 			goto out;
1009 	}
1010 
1011 out:
1012 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
1013 	mlx4_cq_set_ci(&cq->mcq);
1014 	wmb(); /* ensure HW sees CQ consumer before we post new buffers */
1015 	ring->cons = cq->mcq.cons_index;
1016 	mlx4_en_refill_rx_buffers(priv, ring);
1017 	mlx4_en_update_rx_prod_db(ring);
1018 	return polled;
1019 }
1020 
1021 
1022 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
1023 {
1024 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
1025 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
1026 
1027 	if (likely(priv->port_up))
1028 		napi_schedule_irqoff(&cq->napi);
1029 	else
1030 		mlx4_en_arm_cq(priv, cq);
1031 }
1032 
1033 /* Rx CQ polling - called by NAPI */
1034 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
1035 {
1036 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
1037 	struct net_device *dev = cq->dev;
1038 	struct mlx4_en_priv *priv = netdev_priv(dev);
1039 	int done;
1040 
1041 	if (!mlx4_en_cq_lock_napi(cq))
1042 		return budget;
1043 
1044 	done = mlx4_en_process_rx_cq(dev, cq, budget);
1045 
1046 	mlx4_en_cq_unlock_napi(cq);
1047 
1048 	/* If we used up all the quota - we're probably not done yet... */
1049 	if (done == budget) {
1050 		const struct cpumask *aff;
1051 		struct irq_data *idata;
1052 		int cpu_curr;
1053 
1054 		INC_PERF_COUNTER(priv->pstats.napi_quota);
1055 
1056 		cpu_curr = smp_processor_id();
1057 		idata = irq_desc_get_irq_data(cq->irq_desc);
1058 		aff = irq_data_get_affinity_mask(idata);
1059 
1060 		if (likely(cpumask_test_cpu(cpu_curr, aff)))
1061 			return budget;
1062 
1063 		/* Current cpu is not according to smp_irq_affinity -
1064 		 * probably affinity changed. need to stop this NAPI
1065 		 * poll, and restart it on the right CPU
1066 		 */
1067 		done = 0;
1068 	}
1069 	/* Done for now */
1070 	napi_complete_done(napi, done);
1071 	mlx4_en_arm_cq(priv, cq);
1072 	return done;
1073 }
1074 
1075 static const int frag_sizes[] = {
1076 	FRAG_SZ0,
1077 	FRAG_SZ1,
1078 	FRAG_SZ2,
1079 	FRAG_SZ3
1080 };
1081 
1082 void mlx4_en_calc_rx_buf(struct net_device *dev)
1083 {
1084 	struct mlx4_en_priv *priv = netdev_priv(dev);
1085 	/* VLAN_HLEN is added twice,to support skb vlan tagged with multiple
1086 	 * headers. (For example: ETH_P_8021Q and ETH_P_8021AD).
1087 	 */
1088 	int eff_mtu = dev->mtu + ETH_HLEN + (2 * VLAN_HLEN);
1089 	int buf_size = 0;
1090 	int i = 0;
1091 
1092 	while (buf_size < eff_mtu) {
1093 		priv->frag_info[i].frag_size =
1094 			(eff_mtu > buf_size + frag_sizes[i]) ?
1095 				frag_sizes[i] : eff_mtu - buf_size;
1096 		priv->frag_info[i].frag_prefix_size = buf_size;
1097 		priv->frag_info[i].frag_stride =
1098 				ALIGN(priv->frag_info[i].frag_size,
1099 				      SMP_CACHE_BYTES);
1100 		buf_size += priv->frag_info[i].frag_size;
1101 		i++;
1102 	}
1103 
1104 	priv->num_frags = i;
1105 	priv->rx_skb_size = eff_mtu;
1106 	priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
1107 
1108 	en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d num_frags:%d):\n",
1109 	       eff_mtu, priv->num_frags);
1110 	for (i = 0; i < priv->num_frags; i++) {
1111 		en_err(priv,
1112 		       "  frag:%d - size:%d prefix:%d stride:%d\n",
1113 		       i,
1114 		       priv->frag_info[i].frag_size,
1115 		       priv->frag_info[i].frag_prefix_size,
1116 		       priv->frag_info[i].frag_stride);
1117 	}
1118 }
1119 
1120 /* RSS related functions */
1121 
1122 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
1123 				 struct mlx4_en_rx_ring *ring,
1124 				 enum mlx4_qp_state *state,
1125 				 struct mlx4_qp *qp)
1126 {
1127 	struct mlx4_en_dev *mdev = priv->mdev;
1128 	struct mlx4_qp_context *context;
1129 	int err = 0;
1130 
1131 	context = kmalloc(sizeof(*context), GFP_KERNEL);
1132 	if (!context)
1133 		return -ENOMEM;
1134 
1135 	err = mlx4_qp_alloc(mdev->dev, qpn, qp, GFP_KERNEL);
1136 	if (err) {
1137 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
1138 		goto out;
1139 	}
1140 	qp->event = mlx4_en_sqp_event;
1141 
1142 	memset(context, 0, sizeof *context);
1143 	mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
1144 				qpn, ring->cqn, -1, context);
1145 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
1146 
1147 	/* Cancel FCS removal if FW allows */
1148 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
1149 		context->param3 |= cpu_to_be32(1 << 29);
1150 		if (priv->dev->features & NETIF_F_RXFCS)
1151 			ring->fcs_del = 0;
1152 		else
1153 			ring->fcs_del = ETH_FCS_LEN;
1154 	} else
1155 		ring->fcs_del = 0;
1156 
1157 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
1158 	if (err) {
1159 		mlx4_qp_remove(mdev->dev, qp);
1160 		mlx4_qp_free(mdev->dev, qp);
1161 	}
1162 	mlx4_en_update_rx_prod_db(ring);
1163 out:
1164 	kfree(context);
1165 	return err;
1166 }
1167 
1168 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
1169 {
1170 	int err;
1171 	u32 qpn;
1172 
1173 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn,
1174 				    MLX4_RESERVE_A0_QP);
1175 	if (err) {
1176 		en_err(priv, "Failed reserving drop qpn\n");
1177 		return err;
1178 	}
1179 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp, GFP_KERNEL);
1180 	if (err) {
1181 		en_err(priv, "Failed allocating drop qp\n");
1182 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1183 		return err;
1184 	}
1185 
1186 	return 0;
1187 }
1188 
1189 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
1190 {
1191 	u32 qpn;
1192 
1193 	qpn = priv->drop_qp.qpn;
1194 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1195 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1196 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1197 }
1198 
1199 /* Allocate rx qp's and configure them according to rss map */
1200 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1201 {
1202 	struct mlx4_en_dev *mdev = priv->mdev;
1203 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1204 	struct mlx4_qp_context context;
1205 	struct mlx4_rss_context *rss_context;
1206 	int rss_rings;
1207 	void *ptr;
1208 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1209 			MLX4_RSS_TCP_IPV6);
1210 	int i, qpn;
1211 	int err = 0;
1212 	int good_qps = 0;
1213 
1214 	en_dbg(DRV, priv, "Configuring rss steering\n");
1215 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1216 				    priv->rx_ring_num,
1217 				    &rss_map->base_qpn, 0);
1218 	if (err) {
1219 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1220 		return err;
1221 	}
1222 
1223 	for (i = 0; i < priv->rx_ring_num; i++) {
1224 		qpn = rss_map->base_qpn + i;
1225 		err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1226 					    &rss_map->state[i],
1227 					    &rss_map->qps[i]);
1228 		if (err)
1229 			goto rss_err;
1230 
1231 		++good_qps;
1232 	}
1233 
1234 	/* Configure RSS indirection qp */
1235 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp, GFP_KERNEL);
1236 	if (err) {
1237 		en_err(priv, "Failed to allocate RSS indirection QP\n");
1238 		goto rss_err;
1239 	}
1240 	rss_map->indir_qp.event = mlx4_en_sqp_event;
1241 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1242 				priv->rx_ring[0]->cqn, -1, &context);
1243 
1244 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1245 		rss_rings = priv->rx_ring_num;
1246 	else
1247 		rss_rings = priv->prof->rss_rings;
1248 
1249 	ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1250 					+ MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1251 	rss_context = ptr;
1252 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1253 					    (rss_map->base_qpn));
1254 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1255 	if (priv->mdev->profile.udp_rss) {
1256 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1257 		rss_context->base_qpn_udp = rss_context->default_qpn;
1258 	}
1259 
1260 	if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1261 		en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1262 		rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1263 	}
1264 
1265 	rss_context->flags = rss_mask;
1266 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1267 	if (priv->rss_hash_fn == ETH_RSS_HASH_XOR) {
1268 		rss_context->hash_fn = MLX4_RSS_HASH_XOR;
1269 	} else if (priv->rss_hash_fn == ETH_RSS_HASH_TOP) {
1270 		rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1271 		memcpy(rss_context->rss_key, priv->rss_key,
1272 		       MLX4_EN_RSS_KEY_SIZE);
1273 	} else {
1274 		en_err(priv, "Unknown RSS hash function requested\n");
1275 		err = -EINVAL;
1276 		goto indir_err;
1277 	}
1278 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1279 			       &rss_map->indir_qp, &rss_map->indir_state);
1280 	if (err)
1281 		goto indir_err;
1282 
1283 	return 0;
1284 
1285 indir_err:
1286 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1287 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1288 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1289 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1290 rss_err:
1291 	for (i = 0; i < good_qps; i++) {
1292 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1293 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1294 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1295 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1296 	}
1297 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1298 	return err;
1299 }
1300 
1301 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1302 {
1303 	struct mlx4_en_dev *mdev = priv->mdev;
1304 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1305 	int i;
1306 
1307 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1308 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1309 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1310 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1311 
1312 	for (i = 0; i < priv->rx_ring_num; i++) {
1313 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1314 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1315 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1316 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1317 	}
1318 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1319 }
1320