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/bpf.h>
36 #include <linux/mlx4/cq.h>
37 #include <linux/slab.h>
38 #include <linux/mlx4/qp.h>
39 #include <linux/skbuff.h>
40 #include <linux/rculist.h>
41 #include <linux/if_ether.h>
42 #include <linux/if_vlan.h>
43 #include <linux/vmalloc.h>
44 #include <linux/irq.h>
45 
46 #if IS_ENABLED(CONFIG_IPV6)
47 #include <net/ip6_checksum.h>
48 #endif
49 
50 #include "mlx4_en.h"
51 
52 static int mlx4_alloc_pages(struct mlx4_en_priv *priv,
53 			    struct mlx4_en_rx_alloc *page_alloc,
54 			    const struct mlx4_en_frag_info *frag_info,
55 			    gfp_t _gfp)
56 {
57 	int order;
58 	struct page *page;
59 	dma_addr_t dma;
60 
61 	for (order = frag_info->order; ;) {
62 		gfp_t gfp = _gfp;
63 
64 		if (order)
65 			gfp |= __GFP_COMP | __GFP_NOWARN | __GFP_NOMEMALLOC;
66 		page = alloc_pages(gfp, order);
67 		if (likely(page))
68 			break;
69 		if (--order < 0 ||
70 		    ((PAGE_SIZE << order) < frag_info->frag_size))
71 			return -ENOMEM;
72 	}
73 	dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE << order,
74 			   frag_info->dma_dir);
75 	if (dma_mapping_error(priv->ddev, dma)) {
76 		put_page(page);
77 		return -ENOMEM;
78 	}
79 	page_alloc->page_size = PAGE_SIZE << order;
80 	page_alloc->page = page;
81 	page_alloc->dma = dma;
82 	page_alloc->page_offset = 0;
83 	/* Not doing get_page() for each frag is a big win
84 	 * on asymetric workloads. Note we can not use atomic_set().
85 	 */
86 	page_ref_add(page, page_alloc->page_size / frag_info->frag_stride - 1);
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,
129 				priv->frag_info[i].dma_dir);
130 			page = page_alloc[i].page;
131 			/* Revert changes done by mlx4_alloc_pages */
132 			page_ref_sub(page, page_alloc[i].page_size /
133 					   priv->frag_info[i].frag_stride - 1);
134 			put_page(page);
135 		}
136 	}
137 	return -ENOMEM;
138 }
139 
140 static void mlx4_en_free_frag(struct mlx4_en_priv *priv,
141 			      struct mlx4_en_rx_alloc *frags,
142 			      int i)
143 {
144 	const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
145 	u32 next_frag_end = frags[i].page_offset + 2 * frag_info->frag_stride;
146 
147 
148 	if (next_frag_end > frags[i].page_size)
149 		dma_unmap_page(priv->ddev, frags[i].dma, frags[i].page_size,
150 			       frag_info->dma_dir);
151 
152 	if (frags[i].page)
153 		put_page(frags[i].page);
154 }
155 
156 static int mlx4_en_init_allocator(struct mlx4_en_priv *priv,
157 				  struct mlx4_en_rx_ring *ring)
158 {
159 	int i;
160 	struct mlx4_en_rx_alloc *page_alloc;
161 
162 	for (i = 0; i < priv->num_frags; i++) {
163 		const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
164 
165 		if (mlx4_alloc_pages(priv, &ring->page_alloc[i],
166 				     frag_info, GFP_KERNEL | __GFP_COLD))
167 			goto out;
168 
169 		en_dbg(DRV, priv, "  frag %d allocator: - size:%d frags:%d\n",
170 		       i, ring->page_alloc[i].page_size,
171 		       page_ref_count(ring->page_alloc[i].page));
172 	}
173 	return 0;
174 
175 out:
176 	while (i--) {
177 		struct page *page;
178 
179 		page_alloc = &ring->page_alloc[i];
180 		dma_unmap_page(priv->ddev, page_alloc->dma,
181 			       page_alloc->page_size,
182 			       priv->frag_info[i].dma_dir);
183 		page = page_alloc->page;
184 		/* Revert changes done by mlx4_alloc_pages */
185 		page_ref_sub(page, page_alloc->page_size /
186 				   priv->frag_info[i].frag_stride - 1);
187 		put_page(page);
188 		page_alloc->page = NULL;
189 	}
190 	return -ENOMEM;
191 }
192 
193 static void mlx4_en_destroy_allocator(struct mlx4_en_priv *priv,
194 				      struct mlx4_en_rx_ring *ring)
195 {
196 	struct mlx4_en_rx_alloc *page_alloc;
197 	int i;
198 
199 	for (i = 0; i < priv->num_frags; i++) {
200 		const struct mlx4_en_frag_info *frag_info = &priv->frag_info[i];
201 
202 		page_alloc = &ring->page_alloc[i];
203 		en_dbg(DRV, priv, "Freeing allocator:%d count:%d\n",
204 		       i, page_count(page_alloc->page));
205 
206 		dma_unmap_page(priv->ddev, page_alloc->dma,
207 				page_alloc->page_size, frag_info->dma_dir);
208 		while (page_alloc->page_offset + frag_info->frag_stride <
209 		       page_alloc->page_size) {
210 			put_page(page_alloc->page);
211 			page_alloc->page_offset += frag_info->frag_stride;
212 		}
213 		page_alloc->page = NULL;
214 	}
215 }
216 
217 static void mlx4_en_init_rx_desc(struct mlx4_en_priv *priv,
218 				 struct mlx4_en_rx_ring *ring, int index)
219 {
220 	struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
221 	int possible_frags;
222 	int i;
223 
224 	/* Set size and memtype fields */
225 	for (i = 0; i < priv->num_frags; i++) {
226 		rx_desc->data[i].byte_count =
227 			cpu_to_be32(priv->frag_info[i].frag_size);
228 		rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
229 	}
230 
231 	/* If the number of used fragments does not fill up the ring stride,
232 	 * remaining (unused) fragments must be padded with null address/size
233 	 * and a special memory key */
234 	possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
235 	for (i = priv->num_frags; i < possible_frags; i++) {
236 		rx_desc->data[i].byte_count = 0;
237 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
238 		rx_desc->data[i].addr = 0;
239 	}
240 }
241 
242 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
243 				   struct mlx4_en_rx_ring *ring, int index,
244 				   gfp_t gfp)
245 {
246 	struct mlx4_en_rx_desc *rx_desc = ring->buf + (index * ring->stride);
247 	struct mlx4_en_rx_alloc *frags = ring->rx_info +
248 					(index << priv->log_rx_info);
249 
250 	if (ring->page_cache.index > 0) {
251 		frags[0] = ring->page_cache.buf[--ring->page_cache.index];
252 		rx_desc->data[0].addr = cpu_to_be64(frags[0].dma);
253 		return 0;
254 	}
255 
256 	return mlx4_en_alloc_frags(priv, rx_desc, frags, ring->page_alloc, gfp);
257 }
258 
259 static inline bool mlx4_en_is_ring_empty(struct mlx4_en_rx_ring *ring)
260 {
261 	return ring->prod == ring->cons;
262 }
263 
264 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
265 {
266 	*ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
267 }
268 
269 static void mlx4_en_free_rx_desc(struct mlx4_en_priv *priv,
270 				 struct mlx4_en_rx_ring *ring,
271 				 int index)
272 {
273 	struct mlx4_en_rx_alloc *frags;
274 	int nr;
275 
276 	frags = ring->rx_info + (index << priv->log_rx_info);
277 	for (nr = 0; nr < priv->num_frags; nr++) {
278 		en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
279 		mlx4_en_free_frag(priv, frags, nr);
280 	}
281 }
282 
283 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
284 {
285 	struct mlx4_en_rx_ring *ring;
286 	int ring_ind;
287 	int buf_ind;
288 	int new_size;
289 
290 	for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
291 		for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
292 			ring = priv->rx_ring[ring_ind];
293 
294 			if (mlx4_en_prepare_rx_desc(priv, ring,
295 						    ring->actual_size,
296 						    GFP_KERNEL | __GFP_COLD)) {
297 				if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
298 					en_err(priv, "Failed to allocate enough rx buffers\n");
299 					return -ENOMEM;
300 				} else {
301 					new_size = rounddown_pow_of_two(ring->actual_size);
302 					en_warn(priv, "Only %d buffers allocated reducing ring size to %d\n",
303 						ring->actual_size, new_size);
304 					goto reduce_rings;
305 				}
306 			}
307 			ring->actual_size++;
308 			ring->prod++;
309 		}
310 	}
311 	return 0;
312 
313 reduce_rings:
314 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
315 		ring = priv->rx_ring[ring_ind];
316 		while (ring->actual_size > new_size) {
317 			ring->actual_size--;
318 			ring->prod--;
319 			mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
320 		}
321 	}
322 
323 	return 0;
324 }
325 
326 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
327 				struct mlx4_en_rx_ring *ring)
328 {
329 	int index;
330 
331 	en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
332 	       ring->cons, ring->prod);
333 
334 	/* Unmap and free Rx buffers */
335 	while (!mlx4_en_is_ring_empty(ring)) {
336 		index = ring->cons & ring->size_mask;
337 		en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
338 		mlx4_en_free_rx_desc(priv, ring, index);
339 		++ring->cons;
340 	}
341 }
342 
343 void mlx4_en_set_num_rx_rings(struct mlx4_en_dev *mdev)
344 {
345 	int i;
346 	int num_of_eqs;
347 	int num_rx_rings;
348 	struct mlx4_dev *dev = mdev->dev;
349 
350 	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH) {
351 		num_of_eqs = max_t(int, MIN_RX_RINGS,
352 				   min_t(int,
353 					 mlx4_get_eqs_per_port(mdev->dev, i),
354 					 DEF_RX_RINGS));
355 
356 		num_rx_rings = mlx4_low_memory_profile() ? MIN_RX_RINGS :
357 			min_t(int, num_of_eqs,
358 			      netif_get_num_default_rss_queues());
359 		mdev->profile.prof[i].rx_ring_num =
360 			rounddown_pow_of_two(num_rx_rings);
361 	}
362 }
363 
364 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
365 			   struct mlx4_en_rx_ring **pring,
366 			   u32 size, u16 stride, int node)
367 {
368 	struct mlx4_en_dev *mdev = priv->mdev;
369 	struct mlx4_en_rx_ring *ring;
370 	int err = -ENOMEM;
371 	int tmp;
372 
373 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
374 	if (!ring) {
375 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
376 		if (!ring) {
377 			en_err(priv, "Failed to allocate RX ring structure\n");
378 			return -ENOMEM;
379 		}
380 	}
381 
382 	ring->prod = 0;
383 	ring->cons = 0;
384 	ring->size = size;
385 	ring->size_mask = size - 1;
386 	ring->stride = stride;
387 	ring->log_stride = ffs(ring->stride) - 1;
388 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
389 
390 	tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
391 					sizeof(struct mlx4_en_rx_alloc));
392 	ring->rx_info = vmalloc_node(tmp, node);
393 	if (!ring->rx_info) {
394 		ring->rx_info = vmalloc(tmp);
395 		if (!ring->rx_info) {
396 			err = -ENOMEM;
397 			goto err_ring;
398 		}
399 	}
400 
401 	en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
402 		 ring->rx_info, tmp);
403 
404 	/* Allocate HW buffers on provided NUMA node */
405 	set_dev_node(&mdev->dev->persist->pdev->dev, node);
406 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
407 	set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
408 	if (err)
409 		goto err_info;
410 
411 	ring->buf = ring->wqres.buf.direct.buf;
412 
413 	ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
414 
415 	*pring = ring;
416 	return 0;
417 
418 err_info:
419 	vfree(ring->rx_info);
420 	ring->rx_info = NULL;
421 err_ring:
422 	kfree(ring);
423 	*pring = NULL;
424 
425 	return err;
426 }
427 
428 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
429 {
430 	struct mlx4_en_rx_ring *ring;
431 	int i;
432 	int ring_ind;
433 	int err;
434 	int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
435 					DS_SIZE * priv->num_frags);
436 
437 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
438 		ring = priv->rx_ring[ring_ind];
439 
440 		ring->prod = 0;
441 		ring->cons = 0;
442 		ring->actual_size = 0;
443 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
444 
445 		ring->stride = stride;
446 		if (ring->stride <= TXBB_SIZE)
447 			ring->buf += TXBB_SIZE;
448 
449 		ring->log_stride = ffs(ring->stride) - 1;
450 		ring->buf_size = ring->size * ring->stride;
451 
452 		memset(ring->buf, 0, ring->buf_size);
453 		mlx4_en_update_rx_prod_db(ring);
454 
455 		/* Initialize all descriptors */
456 		for (i = 0; i < ring->size; i++)
457 			mlx4_en_init_rx_desc(priv, ring, i);
458 
459 		/* Initialize page allocators */
460 		err = mlx4_en_init_allocator(priv, ring);
461 		if (err) {
462 			en_err(priv, "Failed initializing ring allocator\n");
463 			if (ring->stride <= TXBB_SIZE)
464 				ring->buf -= TXBB_SIZE;
465 			ring_ind--;
466 			goto err_allocator;
467 		}
468 	}
469 	err = mlx4_en_fill_rx_buffers(priv);
470 	if (err)
471 		goto err_buffers;
472 
473 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
474 		ring = priv->rx_ring[ring_ind];
475 
476 		ring->size_mask = ring->actual_size - 1;
477 		mlx4_en_update_rx_prod_db(ring);
478 	}
479 
480 	return 0;
481 
482 err_buffers:
483 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
484 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
485 
486 	ring_ind = priv->rx_ring_num - 1;
487 err_allocator:
488 	while (ring_ind >= 0) {
489 		if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
490 			priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
491 		mlx4_en_destroy_allocator(priv, priv->rx_ring[ring_ind]);
492 		ring_ind--;
493 	}
494 	return err;
495 }
496 
497 /* We recover from out of memory by scheduling our napi poll
498  * function (mlx4_en_process_cq), which tries to allocate
499  * all missing RX buffers (call to mlx4_en_refill_rx_buffers).
500  */
501 void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv)
502 {
503 	int ring;
504 
505 	if (!priv->port_up)
506 		return;
507 
508 	for (ring = 0; ring < priv->rx_ring_num; ring++) {
509 		if (mlx4_en_is_ring_empty(priv->rx_ring[ring]))
510 			napi_reschedule(&priv->rx_cq[ring]->napi);
511 	}
512 }
513 
514 /* When the rx ring is running in page-per-packet mode, a released frame can go
515  * directly into a small cache, to avoid unmapping or touching the page
516  * allocator. In bpf prog performance scenarios, buffers are either forwarded
517  * or dropped, never converted to skbs, so every page can come directly from
518  * this cache when it is sized to be a multiple of the napi budget.
519  */
520 bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
521 			struct mlx4_en_rx_alloc *frame)
522 {
523 	struct mlx4_en_page_cache *cache = &ring->page_cache;
524 
525 	if (cache->index >= MLX4_EN_CACHE_SIZE)
526 		return false;
527 
528 	cache->buf[cache->index++] = *frame;
529 	return true;
530 }
531 
532 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
533 			     struct mlx4_en_rx_ring **pring,
534 			     u32 size, u16 stride)
535 {
536 	struct mlx4_en_dev *mdev = priv->mdev;
537 	struct mlx4_en_rx_ring *ring = *pring;
538 	struct bpf_prog *old_prog;
539 
540 	old_prog = READ_ONCE(ring->xdp_prog);
541 	if (old_prog)
542 		bpf_prog_put(old_prog);
543 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
544 	vfree(ring->rx_info);
545 	ring->rx_info = NULL;
546 	kfree(ring);
547 	*pring = NULL;
548 }
549 
550 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
551 				struct mlx4_en_rx_ring *ring)
552 {
553 	int i;
554 
555 	for (i = 0; i < ring->page_cache.index; i++) {
556 		struct mlx4_en_rx_alloc *frame = &ring->page_cache.buf[i];
557 
558 		dma_unmap_page(priv->ddev, frame->dma, frame->page_size,
559 			       priv->frag_info[0].dma_dir);
560 		put_page(frame->page);
561 	}
562 	ring->page_cache.index = 0;
563 	mlx4_en_free_rx_buf(priv, ring);
564 	if (ring->stride <= TXBB_SIZE)
565 		ring->buf -= TXBB_SIZE;
566 	mlx4_en_destroy_allocator(priv, ring);
567 }
568 
569 
570 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
571 				    struct mlx4_en_rx_desc *rx_desc,
572 				    struct mlx4_en_rx_alloc *frags,
573 				    struct sk_buff *skb,
574 				    int length)
575 {
576 	struct skb_frag_struct *skb_frags_rx = skb_shinfo(skb)->frags;
577 	struct mlx4_en_frag_info *frag_info;
578 	int nr;
579 	dma_addr_t dma;
580 
581 	/* Collect used fragments while replacing them in the HW descriptors */
582 	for (nr = 0; nr < priv->num_frags; nr++) {
583 		frag_info = &priv->frag_info[nr];
584 		if (length <= frag_info->frag_prefix_size)
585 			break;
586 		if (!frags[nr].page)
587 			goto fail;
588 
589 		dma = be64_to_cpu(rx_desc->data[nr].addr);
590 		dma_sync_single_for_cpu(priv->ddev, dma, frag_info->frag_size,
591 					DMA_FROM_DEVICE);
592 
593 		/* Save page reference in skb */
594 		__skb_frag_set_page(&skb_frags_rx[nr], frags[nr].page);
595 		skb_frag_size_set(&skb_frags_rx[nr], frag_info->frag_size);
596 		skb_frags_rx[nr].page_offset = frags[nr].page_offset;
597 		skb->truesize += frag_info->frag_stride;
598 		frags[nr].page = NULL;
599 	}
600 	/* Adjust size of last fragment to match actual length */
601 	if (nr > 0)
602 		skb_frag_size_set(&skb_frags_rx[nr - 1],
603 			length - priv->frag_info[nr - 1].frag_prefix_size);
604 	return nr;
605 
606 fail:
607 	while (nr > 0) {
608 		nr--;
609 		__skb_frag_unref(&skb_frags_rx[nr]);
610 	}
611 	return 0;
612 }
613 
614 
615 static struct sk_buff *mlx4_en_rx_skb(struct mlx4_en_priv *priv,
616 				      struct mlx4_en_rx_desc *rx_desc,
617 				      struct mlx4_en_rx_alloc *frags,
618 				      unsigned int length)
619 {
620 	struct sk_buff *skb;
621 	void *va;
622 	int used_frags;
623 	dma_addr_t dma;
624 
625 	skb = netdev_alloc_skb(priv->dev, SMALL_PACKET_SIZE + NET_IP_ALIGN);
626 	if (!skb) {
627 		en_dbg(RX_ERR, priv, "Failed allocating skb\n");
628 		return NULL;
629 	}
630 	skb_reserve(skb, NET_IP_ALIGN);
631 	skb->len = length;
632 
633 	/* Get pointer to first fragment so we could copy the headers into the
634 	 * (linear part of the) skb */
635 	va = page_address(frags[0].page) + frags[0].page_offset;
636 
637 	if (length <= SMALL_PACKET_SIZE) {
638 		/* We are copying all relevant data to the skb - temporarily
639 		 * sync buffers for the copy */
640 		dma = be64_to_cpu(rx_desc->data[0].addr);
641 		dma_sync_single_for_cpu(priv->ddev, dma, length,
642 					DMA_FROM_DEVICE);
643 		skb_copy_to_linear_data(skb, va, length);
644 		skb->tail += length;
645 	} else {
646 		unsigned int pull_len;
647 
648 		/* Move relevant fragments to skb */
649 		used_frags = mlx4_en_complete_rx_desc(priv, rx_desc, frags,
650 							skb, length);
651 		if (unlikely(!used_frags)) {
652 			kfree_skb(skb);
653 			return NULL;
654 		}
655 		skb_shinfo(skb)->nr_frags = used_frags;
656 
657 		pull_len = eth_get_headlen(va, SMALL_PACKET_SIZE);
658 		/* Copy headers into the skb linear buffer */
659 		memcpy(skb->data, va, pull_len);
660 		skb->tail += pull_len;
661 
662 		/* Skip headers in first fragment */
663 		skb_shinfo(skb)->frags[0].page_offset += pull_len;
664 
665 		/* Adjust size of first fragment */
666 		skb_frag_size_sub(&skb_shinfo(skb)->frags[0], pull_len);
667 		skb->data_len = length - pull_len;
668 	}
669 	return skb;
670 }
671 
672 static void validate_loopback(struct mlx4_en_priv *priv, struct sk_buff *skb)
673 {
674 	int i;
675 	int offset = ETH_HLEN;
676 
677 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++, offset++) {
678 		if (*(skb->data + offset) != (unsigned char) (i & 0xff))
679 			goto out_loopback;
680 	}
681 	/* Loopback found */
682 	priv->loopback_ok = 1;
683 
684 out_loopback:
685 	dev_kfree_skb_any(skb);
686 }
687 
688 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
689 				     struct mlx4_en_rx_ring *ring)
690 {
691 	int index = ring->prod & ring->size_mask;
692 
693 	while ((u32) (ring->prod - ring->cons) < ring->actual_size) {
694 		if (mlx4_en_prepare_rx_desc(priv, ring, index,
695 					    GFP_ATOMIC | __GFP_COLD))
696 			break;
697 		ring->prod++;
698 		index = ring->prod & ring->size_mask;
699 	}
700 }
701 
702 /* When hardware doesn't strip the vlan, we need to calculate the checksum
703  * over it and add it to the hardware's checksum calculation
704  */
705 static inline __wsum get_fixed_vlan_csum(__wsum hw_checksum,
706 					 struct vlan_hdr *vlanh)
707 {
708 	return csum_add(hw_checksum, *(__wsum *)vlanh);
709 }
710 
711 /* Although the stack expects checksum which doesn't include the pseudo
712  * header, the HW adds it. To address that, we are subtracting the pseudo
713  * header checksum from the checksum value provided by the HW.
714  */
715 static void get_fixed_ipv4_csum(__wsum hw_checksum, struct sk_buff *skb,
716 				struct iphdr *iph)
717 {
718 	__u16 length_for_csum = 0;
719 	__wsum csum_pseudo_header = 0;
720 
721 	length_for_csum = (be16_to_cpu(iph->tot_len) - (iph->ihl << 2));
722 	csum_pseudo_header = csum_tcpudp_nofold(iph->saddr, iph->daddr,
723 						length_for_csum, iph->protocol, 0);
724 	skb->csum = csum_sub(hw_checksum, csum_pseudo_header);
725 }
726 
727 #if IS_ENABLED(CONFIG_IPV6)
728 /* In IPv6 packets, besides subtracting the pseudo header checksum,
729  * we also compute/add the IP header checksum which
730  * is not added by the HW.
731  */
732 static int get_fixed_ipv6_csum(__wsum hw_checksum, struct sk_buff *skb,
733 			       struct ipv6hdr *ipv6h)
734 {
735 	__wsum csum_pseudo_hdr = 0;
736 
737 	if (ipv6h->nexthdr == IPPROTO_FRAGMENT || ipv6h->nexthdr == IPPROTO_HOPOPTS)
738 		return -1;
739 	hw_checksum = csum_add(hw_checksum, (__force __wsum)htons(ipv6h->nexthdr));
740 
741 	csum_pseudo_hdr = csum_partial(&ipv6h->saddr,
742 				       sizeof(ipv6h->saddr) + sizeof(ipv6h->daddr), 0);
743 	csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ipv6h->payload_len);
744 	csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ntohs(ipv6h->nexthdr));
745 
746 	skb->csum = csum_sub(hw_checksum, csum_pseudo_hdr);
747 	skb->csum = csum_add(skb->csum, csum_partial(ipv6h, sizeof(struct ipv6hdr), 0));
748 	return 0;
749 }
750 #endif
751 static int check_csum(struct mlx4_cqe *cqe, struct sk_buff *skb, void *va,
752 		      netdev_features_t dev_features)
753 {
754 	__wsum hw_checksum = 0;
755 
756 	void *hdr = (u8 *)va + sizeof(struct ethhdr);
757 
758 	hw_checksum = csum_unfold((__force __sum16)cqe->checksum);
759 
760 	if (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK) &&
761 	    !(dev_features & NETIF_F_HW_VLAN_CTAG_RX)) {
762 		hw_checksum = get_fixed_vlan_csum(hw_checksum, hdr);
763 		hdr += sizeof(struct vlan_hdr);
764 	}
765 
766 	if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4))
767 		get_fixed_ipv4_csum(hw_checksum, skb, hdr);
768 #if IS_ENABLED(CONFIG_IPV6)
769 	else if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV6))
770 		if (get_fixed_ipv6_csum(hw_checksum, skb, hdr))
771 			return -1;
772 #endif
773 	return 0;
774 }
775 
776 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
777 {
778 	struct mlx4_en_priv *priv = netdev_priv(dev);
779 	struct mlx4_en_dev *mdev = priv->mdev;
780 	struct mlx4_cqe *cqe;
781 	struct mlx4_en_rx_ring *ring = priv->rx_ring[cq->ring];
782 	struct mlx4_en_rx_alloc *frags;
783 	struct mlx4_en_rx_desc *rx_desc;
784 	struct bpf_prog *xdp_prog;
785 	int doorbell_pending;
786 	struct sk_buff *skb;
787 	int tx_index;
788 	int index;
789 	int nr;
790 	unsigned int length;
791 	int polled = 0;
792 	int ip_summed;
793 	int factor = priv->cqe_factor;
794 	u64 timestamp;
795 	bool l2_tunnel;
796 
797 	if (!priv->port_up)
798 		return 0;
799 
800 	if (budget <= 0)
801 		return polled;
802 
803 	xdp_prog = READ_ONCE(ring->xdp_prog);
804 	doorbell_pending = 0;
805 	tx_index = (priv->tx_ring_num - priv->xdp_ring_num) + cq->ring;
806 
807 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
808 	 * descriptor offset can be deduced from the CQE index instead of
809 	 * reading 'cqe->index' */
810 	index = cq->mcq.cons_index & ring->size_mask;
811 	cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
812 
813 	/* Process all completed CQEs */
814 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
815 		    cq->mcq.cons_index & cq->size)) {
816 
817 		frags = ring->rx_info + (index << priv->log_rx_info);
818 		rx_desc = ring->buf + (index << ring->log_stride);
819 
820 		/*
821 		 * make sure we read the CQE after we read the ownership bit
822 		 */
823 		dma_rmb();
824 
825 		/* Drop packet on bad receive or bad checksum */
826 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
827 						MLX4_CQE_OPCODE_ERROR)) {
828 			en_err(priv, "CQE completed in error - vendor syndrom:%d syndrom:%d\n",
829 			       ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
830 			       ((struct mlx4_err_cqe *)cqe)->syndrome);
831 			goto next;
832 		}
833 		if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
834 			en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
835 			goto next;
836 		}
837 
838 		/* Check if we need to drop the packet if SRIOV is not enabled
839 		 * and not performing the selftest or flb disabled
840 		 */
841 		if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
842 			struct ethhdr *ethh;
843 			dma_addr_t dma;
844 			/* Get pointer to first fragment since we haven't
845 			 * skb yet and cast it to ethhdr struct
846 			 */
847 			dma = be64_to_cpu(rx_desc->data[0].addr);
848 			dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
849 						DMA_FROM_DEVICE);
850 			ethh = (struct ethhdr *)(page_address(frags[0].page) +
851 						 frags[0].page_offset);
852 
853 			if (is_multicast_ether_addr(ethh->h_dest)) {
854 				struct mlx4_mac_entry *entry;
855 				struct hlist_head *bucket;
856 				unsigned int mac_hash;
857 
858 				/* Drop the packet, since HW loopback-ed it */
859 				mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
860 				bucket = &priv->mac_hash[mac_hash];
861 				rcu_read_lock();
862 				hlist_for_each_entry_rcu(entry, bucket, hlist) {
863 					if (ether_addr_equal_64bits(entry->mac,
864 								    ethh->h_source)) {
865 						rcu_read_unlock();
866 						goto next;
867 					}
868 				}
869 				rcu_read_unlock();
870 			}
871 		}
872 
873 		/*
874 		 * Packet is OK - process it.
875 		 */
876 		length = be32_to_cpu(cqe->byte_cnt);
877 		length -= ring->fcs_del;
878 		ring->bytes += length;
879 		ring->packets++;
880 		l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
881 			(cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
882 
883 		/* A bpf program gets first chance to drop the packet. It may
884 		 * read bytes but not past the end of the frag.
885 		 */
886 		if (xdp_prog) {
887 			struct xdp_buff xdp;
888 			dma_addr_t dma;
889 			u32 act;
890 
891 			dma = be64_to_cpu(rx_desc->data[0].addr);
892 			dma_sync_single_for_cpu(priv->ddev, dma,
893 						priv->frag_info[0].frag_size,
894 						DMA_FROM_DEVICE);
895 
896 			xdp.data = page_address(frags[0].page) +
897 							frags[0].page_offset;
898 			xdp.data_end = xdp.data + length;
899 
900 			act = bpf_prog_run_xdp(xdp_prog, &xdp);
901 			switch (act) {
902 			case XDP_PASS:
903 				break;
904 			case XDP_TX:
905 				if (!mlx4_en_xmit_frame(frags, dev,
906 							length, tx_index,
907 							&doorbell_pending))
908 					goto consumed;
909 				break;
910 			default:
911 				bpf_warn_invalid_xdp_action(act);
912 			case XDP_ABORTED:
913 			case XDP_DROP:
914 				if (mlx4_en_rx_recycle(ring, frags))
915 					goto consumed;
916 				goto next;
917 			}
918 		}
919 
920 		if (likely(dev->features & NETIF_F_RXCSUM)) {
921 			if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
922 						      MLX4_CQE_STATUS_UDP)) {
923 				if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
924 				    cqe->checksum == cpu_to_be16(0xffff)) {
925 					ip_summed = CHECKSUM_UNNECESSARY;
926 					ring->csum_ok++;
927 				} else {
928 					ip_summed = CHECKSUM_NONE;
929 					ring->csum_none++;
930 				}
931 			} else {
932 				if (priv->flags & MLX4_EN_FLAG_RX_CSUM_NON_TCP_UDP &&
933 				    (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
934 							       MLX4_CQE_STATUS_IPV6))) {
935 					ip_summed = CHECKSUM_COMPLETE;
936 					ring->csum_complete++;
937 				} else {
938 					ip_summed = CHECKSUM_NONE;
939 					ring->csum_none++;
940 				}
941 			}
942 		} else {
943 			ip_summed = CHECKSUM_NONE;
944 			ring->csum_none++;
945 		}
946 
947 		/* This packet is eligible for GRO if it is:
948 		 * - DIX Ethernet (type interpretation)
949 		 * - TCP/IP (v4)
950 		 * - without IP options
951 		 * - not an IP fragment
952 		 */
953 		if (dev->features & NETIF_F_GRO) {
954 			struct sk_buff *gro_skb = napi_get_frags(&cq->napi);
955 			if (!gro_skb)
956 				goto next;
957 
958 			nr = mlx4_en_complete_rx_desc(priv,
959 				rx_desc, frags, gro_skb,
960 				length);
961 			if (!nr)
962 				goto next;
963 
964 			if (ip_summed == CHECKSUM_COMPLETE) {
965 				void *va = skb_frag_address(skb_shinfo(gro_skb)->frags);
966 				if (check_csum(cqe, gro_skb, va,
967 					       dev->features)) {
968 					ip_summed = CHECKSUM_NONE;
969 					ring->csum_none++;
970 					ring->csum_complete--;
971 				}
972 			}
973 
974 			skb_shinfo(gro_skb)->nr_frags = nr;
975 			gro_skb->len = length;
976 			gro_skb->data_len = length;
977 			gro_skb->ip_summed = ip_summed;
978 
979 			if (l2_tunnel && ip_summed == CHECKSUM_UNNECESSARY)
980 				gro_skb->csum_level = 1;
981 
982 			if ((cqe->vlan_my_qpn &
983 			    cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK)) &&
984 			    (dev->features & NETIF_F_HW_VLAN_CTAG_RX)) {
985 				u16 vid = be16_to_cpu(cqe->sl_vid);
986 
987 				__vlan_hwaccel_put_tag(gro_skb, htons(ETH_P_8021Q), vid);
988 			} else if ((be32_to_cpu(cqe->vlan_my_qpn) &
989 				  MLX4_CQE_SVLAN_PRESENT_MASK) &&
990 				 (dev->features & NETIF_F_HW_VLAN_STAG_RX)) {
991 				__vlan_hwaccel_put_tag(gro_skb,
992 						       htons(ETH_P_8021AD),
993 						       be16_to_cpu(cqe->sl_vid));
994 			}
995 
996 			if (dev->features & NETIF_F_RXHASH)
997 				skb_set_hash(gro_skb,
998 					     be32_to_cpu(cqe->immed_rss_invalid),
999 					     (ip_summed == CHECKSUM_UNNECESSARY) ?
1000 						PKT_HASH_TYPE_L4 :
1001 						PKT_HASH_TYPE_L3);
1002 
1003 			skb_record_rx_queue(gro_skb, cq->ring);
1004 
1005 			if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
1006 				timestamp = mlx4_en_get_cqe_ts(cqe);
1007 				mlx4_en_fill_hwtstamps(mdev,
1008 						       skb_hwtstamps(gro_skb),
1009 						       timestamp);
1010 			}
1011 
1012 			napi_gro_frags(&cq->napi);
1013 			goto next;
1014 		}
1015 
1016 		/* GRO not possible, complete processing here */
1017 		skb = mlx4_en_rx_skb(priv, rx_desc, frags, length);
1018 		if (!skb) {
1019 			ring->dropped++;
1020 			goto next;
1021 		}
1022 
1023                 if (unlikely(priv->validate_loopback)) {
1024 			validate_loopback(priv, skb);
1025 			goto next;
1026 		}
1027 
1028 		if (ip_summed == CHECKSUM_COMPLETE) {
1029 			if (check_csum(cqe, skb, skb->data, dev->features)) {
1030 				ip_summed = CHECKSUM_NONE;
1031 				ring->csum_complete--;
1032 				ring->csum_none++;
1033 			}
1034 		}
1035 
1036 		skb->ip_summed = ip_summed;
1037 		skb->protocol = eth_type_trans(skb, dev);
1038 		skb_record_rx_queue(skb, cq->ring);
1039 
1040 		if (l2_tunnel && ip_summed == CHECKSUM_UNNECESSARY)
1041 			skb->csum_level = 1;
1042 
1043 		if (dev->features & NETIF_F_RXHASH)
1044 			skb_set_hash(skb,
1045 				     be32_to_cpu(cqe->immed_rss_invalid),
1046 				     (ip_summed == CHECKSUM_UNNECESSARY) ?
1047 					PKT_HASH_TYPE_L4 :
1048 					PKT_HASH_TYPE_L3);
1049 
1050 		if ((be32_to_cpu(cqe->vlan_my_qpn) &
1051 		    MLX4_CQE_CVLAN_PRESENT_MASK) &&
1052 		    (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
1053 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), be16_to_cpu(cqe->sl_vid));
1054 		else if ((be32_to_cpu(cqe->vlan_my_qpn) &
1055 			  MLX4_CQE_SVLAN_PRESENT_MASK) &&
1056 			 (dev->features & NETIF_F_HW_VLAN_STAG_RX))
1057 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021AD),
1058 					       be16_to_cpu(cqe->sl_vid));
1059 
1060 		if (ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL) {
1061 			timestamp = mlx4_en_get_cqe_ts(cqe);
1062 			mlx4_en_fill_hwtstamps(mdev, skb_hwtstamps(skb),
1063 					       timestamp);
1064 		}
1065 
1066 		napi_gro_receive(&cq->napi, skb);
1067 next:
1068 		for (nr = 0; nr < priv->num_frags; nr++)
1069 			mlx4_en_free_frag(priv, frags, nr);
1070 
1071 consumed:
1072 		++cq->mcq.cons_index;
1073 		index = (cq->mcq.cons_index) & ring->size_mask;
1074 		cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
1075 		if (++polled == budget)
1076 			goto out;
1077 	}
1078 
1079 out:
1080 	if (doorbell_pending)
1081 		mlx4_en_xmit_doorbell(priv->tx_ring[tx_index]);
1082 
1083 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
1084 	mlx4_cq_set_ci(&cq->mcq);
1085 	wmb(); /* ensure HW sees CQ consumer before we post new buffers */
1086 	ring->cons = cq->mcq.cons_index;
1087 	mlx4_en_refill_rx_buffers(priv, ring);
1088 	mlx4_en_update_rx_prod_db(ring);
1089 	return polled;
1090 }
1091 
1092 
1093 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
1094 {
1095 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
1096 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
1097 
1098 	if (likely(priv->port_up))
1099 		napi_schedule_irqoff(&cq->napi);
1100 	else
1101 		mlx4_en_arm_cq(priv, cq);
1102 }
1103 
1104 /* Rx CQ polling - called by NAPI */
1105 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
1106 {
1107 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
1108 	struct net_device *dev = cq->dev;
1109 	struct mlx4_en_priv *priv = netdev_priv(dev);
1110 	int done;
1111 
1112 	done = mlx4_en_process_rx_cq(dev, cq, budget);
1113 
1114 	/* If we used up all the quota - we're probably not done yet... */
1115 	if (done == budget) {
1116 		const struct cpumask *aff;
1117 		struct irq_data *idata;
1118 		int cpu_curr;
1119 
1120 		INC_PERF_COUNTER(priv->pstats.napi_quota);
1121 
1122 		cpu_curr = smp_processor_id();
1123 		idata = irq_desc_get_irq_data(cq->irq_desc);
1124 		aff = irq_data_get_affinity_mask(idata);
1125 
1126 		if (likely(cpumask_test_cpu(cpu_curr, aff)))
1127 			return budget;
1128 
1129 		/* Current cpu is not according to smp_irq_affinity -
1130 		 * probably affinity changed. need to stop this NAPI
1131 		 * poll, and restart it on the right CPU
1132 		 */
1133 		done = 0;
1134 	}
1135 	/* Done for now */
1136 	napi_complete_done(napi, done);
1137 	mlx4_en_arm_cq(priv, cq);
1138 	return done;
1139 }
1140 
1141 static const int frag_sizes[] = {
1142 	FRAG_SZ0,
1143 	FRAG_SZ1,
1144 	FRAG_SZ2,
1145 	FRAG_SZ3
1146 };
1147 
1148 void mlx4_en_calc_rx_buf(struct net_device *dev)
1149 {
1150 	enum dma_data_direction dma_dir = PCI_DMA_FROMDEVICE;
1151 	struct mlx4_en_priv *priv = netdev_priv(dev);
1152 	int eff_mtu = MLX4_EN_EFF_MTU(dev->mtu);
1153 	int order = MLX4_EN_ALLOC_PREFER_ORDER;
1154 	u32 align = SMP_CACHE_BYTES;
1155 	int buf_size = 0;
1156 	int i = 0;
1157 
1158 	/* bpf requires buffers to be set up as 1 packet per page.
1159 	 * This only works when num_frags == 1.
1160 	 */
1161 	if (priv->xdp_ring_num) {
1162 		dma_dir = PCI_DMA_BIDIRECTIONAL;
1163 		/* This will gain efficient xdp frame recycling at the expense
1164 		 * of more costly truesize accounting
1165 		 */
1166 		align = PAGE_SIZE;
1167 		order = 0;
1168 	}
1169 
1170 	while (buf_size < eff_mtu) {
1171 		priv->frag_info[i].order = order;
1172 		priv->frag_info[i].frag_size =
1173 			(eff_mtu > buf_size + frag_sizes[i]) ?
1174 				frag_sizes[i] : eff_mtu - buf_size;
1175 		priv->frag_info[i].frag_prefix_size = buf_size;
1176 		priv->frag_info[i].frag_stride =
1177 				ALIGN(priv->frag_info[i].frag_size, align);
1178 		priv->frag_info[i].dma_dir = dma_dir;
1179 		buf_size += priv->frag_info[i].frag_size;
1180 		i++;
1181 	}
1182 
1183 	priv->num_frags = i;
1184 	priv->rx_skb_size = eff_mtu;
1185 	priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
1186 
1187 	en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d num_frags:%d):\n",
1188 	       eff_mtu, priv->num_frags);
1189 	for (i = 0; i < priv->num_frags; i++) {
1190 		en_err(priv,
1191 		       "  frag:%d - size:%d prefix:%d stride:%d\n",
1192 		       i,
1193 		       priv->frag_info[i].frag_size,
1194 		       priv->frag_info[i].frag_prefix_size,
1195 		       priv->frag_info[i].frag_stride);
1196 	}
1197 }
1198 
1199 /* RSS related functions */
1200 
1201 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
1202 				 struct mlx4_en_rx_ring *ring,
1203 				 enum mlx4_qp_state *state,
1204 				 struct mlx4_qp *qp)
1205 {
1206 	struct mlx4_en_dev *mdev = priv->mdev;
1207 	struct mlx4_qp_context *context;
1208 	int err = 0;
1209 
1210 	context = kmalloc(sizeof(*context), GFP_KERNEL);
1211 	if (!context)
1212 		return -ENOMEM;
1213 
1214 	err = mlx4_qp_alloc(mdev->dev, qpn, qp, GFP_KERNEL);
1215 	if (err) {
1216 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
1217 		goto out;
1218 	}
1219 	qp->event = mlx4_en_sqp_event;
1220 
1221 	memset(context, 0, sizeof *context);
1222 	mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
1223 				qpn, ring->cqn, -1, context);
1224 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
1225 
1226 	/* Cancel FCS removal if FW allows */
1227 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
1228 		context->param3 |= cpu_to_be32(1 << 29);
1229 		if (priv->dev->features & NETIF_F_RXFCS)
1230 			ring->fcs_del = 0;
1231 		else
1232 			ring->fcs_del = ETH_FCS_LEN;
1233 	} else
1234 		ring->fcs_del = 0;
1235 
1236 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
1237 	if (err) {
1238 		mlx4_qp_remove(mdev->dev, qp);
1239 		mlx4_qp_free(mdev->dev, qp);
1240 	}
1241 	mlx4_en_update_rx_prod_db(ring);
1242 out:
1243 	kfree(context);
1244 	return err;
1245 }
1246 
1247 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
1248 {
1249 	int err;
1250 	u32 qpn;
1251 
1252 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn,
1253 				    MLX4_RESERVE_A0_QP);
1254 	if (err) {
1255 		en_err(priv, "Failed reserving drop qpn\n");
1256 		return err;
1257 	}
1258 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp, GFP_KERNEL);
1259 	if (err) {
1260 		en_err(priv, "Failed allocating drop qp\n");
1261 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1262 		return err;
1263 	}
1264 
1265 	return 0;
1266 }
1267 
1268 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
1269 {
1270 	u32 qpn;
1271 
1272 	qpn = priv->drop_qp.qpn;
1273 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1274 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1275 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1276 }
1277 
1278 /* Allocate rx qp's and configure them according to rss map */
1279 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1280 {
1281 	struct mlx4_en_dev *mdev = priv->mdev;
1282 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1283 	struct mlx4_qp_context context;
1284 	struct mlx4_rss_context *rss_context;
1285 	int rss_rings;
1286 	void *ptr;
1287 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1288 			MLX4_RSS_TCP_IPV6);
1289 	int i, qpn;
1290 	int err = 0;
1291 	int good_qps = 0;
1292 
1293 	en_dbg(DRV, priv, "Configuring rss steering\n");
1294 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1295 				    priv->rx_ring_num,
1296 				    &rss_map->base_qpn, 0);
1297 	if (err) {
1298 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1299 		return err;
1300 	}
1301 
1302 	for (i = 0; i < priv->rx_ring_num; i++) {
1303 		qpn = rss_map->base_qpn + i;
1304 		err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1305 					    &rss_map->state[i],
1306 					    &rss_map->qps[i]);
1307 		if (err)
1308 			goto rss_err;
1309 
1310 		++good_qps;
1311 	}
1312 
1313 	/* Configure RSS indirection qp */
1314 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, &rss_map->indir_qp, GFP_KERNEL);
1315 	if (err) {
1316 		en_err(priv, "Failed to allocate RSS indirection QP\n");
1317 		goto rss_err;
1318 	}
1319 	rss_map->indir_qp.event = mlx4_en_sqp_event;
1320 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1321 				priv->rx_ring[0]->cqn, -1, &context);
1322 
1323 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1324 		rss_rings = priv->rx_ring_num;
1325 	else
1326 		rss_rings = priv->prof->rss_rings;
1327 
1328 	ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1329 					+ MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1330 	rss_context = ptr;
1331 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1332 					    (rss_map->base_qpn));
1333 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1334 	if (priv->mdev->profile.udp_rss) {
1335 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1336 		rss_context->base_qpn_udp = rss_context->default_qpn;
1337 	}
1338 
1339 	if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1340 		en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1341 		rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1342 	}
1343 
1344 	rss_context->flags = rss_mask;
1345 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1346 	if (priv->rss_hash_fn == ETH_RSS_HASH_XOR) {
1347 		rss_context->hash_fn = MLX4_RSS_HASH_XOR;
1348 	} else if (priv->rss_hash_fn == ETH_RSS_HASH_TOP) {
1349 		rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1350 		memcpy(rss_context->rss_key, priv->rss_key,
1351 		       MLX4_EN_RSS_KEY_SIZE);
1352 	} else {
1353 		en_err(priv, "Unknown RSS hash function requested\n");
1354 		err = -EINVAL;
1355 		goto indir_err;
1356 	}
1357 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1358 			       &rss_map->indir_qp, &rss_map->indir_state);
1359 	if (err)
1360 		goto indir_err;
1361 
1362 	return 0;
1363 
1364 indir_err:
1365 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1366 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1367 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1368 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1369 rss_err:
1370 	for (i = 0; i < good_qps; i++) {
1371 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1372 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1373 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1374 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1375 	}
1376 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1377 	return err;
1378 }
1379 
1380 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1381 {
1382 	struct mlx4_en_dev *mdev = priv->mdev;
1383 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1384 	int i;
1385 
1386 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1387 		       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->indir_qp);
1388 	mlx4_qp_remove(mdev->dev, &rss_map->indir_qp);
1389 	mlx4_qp_free(mdev->dev, &rss_map->indir_qp);
1390 
1391 	for (i = 0; i < priv->rx_ring_num; i++) {
1392 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1393 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1394 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1395 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1396 	}
1397 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1398 }
1399