xref: /openbmc/linux/drivers/net/ethernet/mellanox/mlx4/en_rx.c (revision 4da722ca19f30f7db250db808d1ab1703607a932)
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/bpf_trace.h>
37 #include <linux/mlx4/cq.h>
38 #include <linux/slab.h>
39 #include <linux/mlx4/qp.h>
40 #include <linux/skbuff.h>
41 #include <linux/rculist.h>
42 #include <linux/if_ether.h>
43 #include <linux/if_vlan.h>
44 #include <linux/vmalloc.h>
45 #include <linux/irq.h>
46 
47 #if IS_ENABLED(CONFIG_IPV6)
48 #include <net/ip6_checksum.h>
49 #endif
50 
51 #include "mlx4_en.h"
52 
53 static int mlx4_alloc_page(struct mlx4_en_priv *priv,
54 			   struct mlx4_en_rx_alloc *frag,
55 			   gfp_t gfp)
56 {
57 	struct page *page;
58 	dma_addr_t dma;
59 
60 	page = alloc_page(gfp);
61 	if (unlikely(!page))
62 		return -ENOMEM;
63 	dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE, priv->dma_dir);
64 	if (unlikely(dma_mapping_error(priv->ddev, dma))) {
65 		__free_page(page);
66 		return -ENOMEM;
67 	}
68 	frag->page = page;
69 	frag->dma = dma;
70 	frag->page_offset = priv->rx_headroom;
71 	return 0;
72 }
73 
74 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
75 			       struct mlx4_en_rx_ring *ring,
76 			       struct mlx4_en_rx_desc *rx_desc,
77 			       struct mlx4_en_rx_alloc *frags,
78 			       gfp_t gfp)
79 {
80 	int i;
81 
82 	for (i = 0; i < priv->num_frags; i++, frags++) {
83 		if (!frags->page) {
84 			if (mlx4_alloc_page(priv, frags, gfp))
85 				return -ENOMEM;
86 			ring->rx_alloc_pages++;
87 		}
88 		rx_desc->data[i].addr = cpu_to_be64(frags->dma +
89 						    frags->page_offset);
90 	}
91 	return 0;
92 }
93 
94 static void mlx4_en_free_frag(const struct mlx4_en_priv *priv,
95 			      struct mlx4_en_rx_alloc *frag)
96 {
97 	if (frag->page) {
98 		dma_unmap_page(priv->ddev, frag->dma,
99 			       PAGE_SIZE, priv->dma_dir);
100 		__free_page(frag->page);
101 	}
102 	/* We need to clear all fields, otherwise a change of priv->log_rx_info
103 	 * could lead to see garbage later in frag->page.
104 	 */
105 	memset(frag, 0, sizeof(*frag));
106 }
107 
108 static void mlx4_en_init_rx_desc(const struct mlx4_en_priv *priv,
109 				 struct mlx4_en_rx_ring *ring, int index)
110 {
111 	struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
112 	int possible_frags;
113 	int i;
114 
115 	/* Set size and memtype fields */
116 	for (i = 0; i < priv->num_frags; i++) {
117 		rx_desc->data[i].byte_count =
118 			cpu_to_be32(priv->frag_info[i].frag_size);
119 		rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
120 	}
121 
122 	/* If the number of used fragments does not fill up the ring stride,
123 	 * remaining (unused) fragments must be padded with null address/size
124 	 * and a special memory key */
125 	possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
126 	for (i = priv->num_frags; i < possible_frags; i++) {
127 		rx_desc->data[i].byte_count = 0;
128 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
129 		rx_desc->data[i].addr = 0;
130 	}
131 }
132 
133 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
134 				   struct mlx4_en_rx_ring *ring, int index,
135 				   gfp_t gfp)
136 {
137 	struct mlx4_en_rx_desc *rx_desc = ring->buf +
138 		(index << ring->log_stride);
139 	struct mlx4_en_rx_alloc *frags = ring->rx_info +
140 					(index << priv->log_rx_info);
141 	if (likely(ring->page_cache.index > 0)) {
142 		/* XDP uses a single page per frame */
143 		if (!frags->page) {
144 			ring->page_cache.index--;
145 			frags->page = ring->page_cache.buf[ring->page_cache.index].page;
146 			frags->dma  = ring->page_cache.buf[ring->page_cache.index].dma;
147 		}
148 		frags->page_offset = XDP_PACKET_HEADROOM;
149 		rx_desc->data[0].addr = cpu_to_be64(frags->dma +
150 						    XDP_PACKET_HEADROOM);
151 		return 0;
152 	}
153 
154 	return mlx4_en_alloc_frags(priv, ring, rx_desc, frags, gfp);
155 }
156 
157 static bool mlx4_en_is_ring_empty(const struct mlx4_en_rx_ring *ring)
158 {
159 	return ring->prod == ring->cons;
160 }
161 
162 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
163 {
164 	*ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
165 }
166 
167 /* slow path */
168 static void mlx4_en_free_rx_desc(const struct mlx4_en_priv *priv,
169 				 struct mlx4_en_rx_ring *ring,
170 				 int index)
171 {
172 	struct mlx4_en_rx_alloc *frags;
173 	int nr;
174 
175 	frags = ring->rx_info + (index << priv->log_rx_info);
176 	for (nr = 0; nr < priv->num_frags; nr++) {
177 		en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
178 		mlx4_en_free_frag(priv, frags + nr);
179 	}
180 }
181 
182 /* Function not in fast-path */
183 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
184 {
185 	struct mlx4_en_rx_ring *ring;
186 	int ring_ind;
187 	int buf_ind;
188 	int new_size;
189 
190 	for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
191 		for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
192 			ring = priv->rx_ring[ring_ind];
193 
194 			if (mlx4_en_prepare_rx_desc(priv, ring,
195 						    ring->actual_size,
196 						    GFP_KERNEL | __GFP_COLD)) {
197 				if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
198 					en_err(priv, "Failed to allocate enough rx buffers\n");
199 					return -ENOMEM;
200 				} else {
201 					new_size = rounddown_pow_of_two(ring->actual_size);
202 					en_warn(priv, "Only %d buffers allocated reducing ring size to %d\n",
203 						ring->actual_size, new_size);
204 					goto reduce_rings;
205 				}
206 			}
207 			ring->actual_size++;
208 			ring->prod++;
209 		}
210 	}
211 	return 0;
212 
213 reduce_rings:
214 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
215 		ring = priv->rx_ring[ring_ind];
216 		while (ring->actual_size > new_size) {
217 			ring->actual_size--;
218 			ring->prod--;
219 			mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
220 		}
221 	}
222 
223 	return 0;
224 }
225 
226 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
227 				struct mlx4_en_rx_ring *ring)
228 {
229 	int index;
230 
231 	en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
232 	       ring->cons, ring->prod);
233 
234 	/* Unmap and free Rx buffers */
235 	for (index = 0; index < ring->size; index++) {
236 		en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
237 		mlx4_en_free_rx_desc(priv, ring, index);
238 	}
239 	ring->cons = 0;
240 	ring->prod = 0;
241 }
242 
243 void mlx4_en_set_num_rx_rings(struct mlx4_en_dev *mdev)
244 {
245 	int i;
246 	int num_of_eqs;
247 	int num_rx_rings;
248 	struct mlx4_dev *dev = mdev->dev;
249 
250 	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH) {
251 		num_of_eqs = max_t(int, MIN_RX_RINGS,
252 				   min_t(int,
253 					 mlx4_get_eqs_per_port(mdev->dev, i),
254 					 DEF_RX_RINGS));
255 
256 		num_rx_rings = mlx4_low_memory_profile() ? MIN_RX_RINGS :
257 			min_t(int, num_of_eqs,
258 			      netif_get_num_default_rss_queues());
259 		mdev->profile.prof[i].rx_ring_num =
260 			rounddown_pow_of_two(num_rx_rings);
261 	}
262 }
263 
264 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
265 			   struct mlx4_en_rx_ring **pring,
266 			   u32 size, u16 stride, int node)
267 {
268 	struct mlx4_en_dev *mdev = priv->mdev;
269 	struct mlx4_en_rx_ring *ring;
270 	int err = -ENOMEM;
271 	int tmp;
272 
273 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
274 	if (!ring) {
275 		ring = kzalloc(sizeof(*ring), GFP_KERNEL);
276 		if (!ring) {
277 			en_err(priv, "Failed to allocate RX ring structure\n");
278 			return -ENOMEM;
279 		}
280 	}
281 
282 	ring->prod = 0;
283 	ring->cons = 0;
284 	ring->size = size;
285 	ring->size_mask = size - 1;
286 	ring->stride = stride;
287 	ring->log_stride = ffs(ring->stride) - 1;
288 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
289 
290 	tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
291 					sizeof(struct mlx4_en_rx_alloc));
292 	ring->rx_info = vzalloc_node(tmp, node);
293 	if (!ring->rx_info) {
294 		ring->rx_info = vzalloc(tmp);
295 		if (!ring->rx_info) {
296 			err = -ENOMEM;
297 			goto err_ring;
298 		}
299 	}
300 
301 	en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
302 		 ring->rx_info, tmp);
303 
304 	/* Allocate HW buffers on provided NUMA node */
305 	set_dev_node(&mdev->dev->persist->pdev->dev, node);
306 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
307 	set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
308 	if (err)
309 		goto err_info;
310 
311 	ring->buf = ring->wqres.buf.direct.buf;
312 
313 	ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
314 
315 	*pring = ring;
316 	return 0;
317 
318 err_info:
319 	vfree(ring->rx_info);
320 	ring->rx_info = NULL;
321 err_ring:
322 	kfree(ring);
323 	*pring = NULL;
324 
325 	return err;
326 }
327 
328 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
329 {
330 	struct mlx4_en_rx_ring *ring;
331 	int i;
332 	int ring_ind;
333 	int err;
334 	int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
335 					DS_SIZE * priv->num_frags);
336 
337 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
338 		ring = priv->rx_ring[ring_ind];
339 
340 		ring->prod = 0;
341 		ring->cons = 0;
342 		ring->actual_size = 0;
343 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
344 
345 		ring->stride = stride;
346 		if (ring->stride <= TXBB_SIZE) {
347 			/* Stamp first unused send wqe */
348 			__be32 *ptr = (__be32 *)ring->buf;
349 			__be32 stamp = cpu_to_be32(1 << STAMP_SHIFT);
350 			*ptr = stamp;
351 			/* Move pointer to start of rx section */
352 			ring->buf += TXBB_SIZE;
353 		}
354 
355 		ring->log_stride = ffs(ring->stride) - 1;
356 		ring->buf_size = ring->size * ring->stride;
357 
358 		memset(ring->buf, 0, ring->buf_size);
359 		mlx4_en_update_rx_prod_db(ring);
360 
361 		/* Initialize all descriptors */
362 		for (i = 0; i < ring->size; i++)
363 			mlx4_en_init_rx_desc(priv, ring, i);
364 	}
365 	err = mlx4_en_fill_rx_buffers(priv);
366 	if (err)
367 		goto err_buffers;
368 
369 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
370 		ring = priv->rx_ring[ring_ind];
371 
372 		ring->size_mask = ring->actual_size - 1;
373 		mlx4_en_update_rx_prod_db(ring);
374 	}
375 
376 	return 0;
377 
378 err_buffers:
379 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
380 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
381 
382 	ring_ind = priv->rx_ring_num - 1;
383 	while (ring_ind >= 0) {
384 		if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
385 			priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
386 		ring_ind--;
387 	}
388 	return err;
389 }
390 
391 /* We recover from out of memory by scheduling our napi poll
392  * function (mlx4_en_process_cq), which tries to allocate
393  * all missing RX buffers (call to mlx4_en_refill_rx_buffers).
394  */
395 void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv)
396 {
397 	int ring;
398 
399 	if (!priv->port_up)
400 		return;
401 
402 	for (ring = 0; ring < priv->rx_ring_num; ring++) {
403 		if (mlx4_en_is_ring_empty(priv->rx_ring[ring])) {
404 			local_bh_disable();
405 			napi_reschedule(&priv->rx_cq[ring]->napi);
406 			local_bh_enable();
407 		}
408 	}
409 }
410 
411 /* When the rx ring is running in page-per-packet mode, a released frame can go
412  * directly into a small cache, to avoid unmapping or touching the page
413  * allocator. In bpf prog performance scenarios, buffers are either forwarded
414  * or dropped, never converted to skbs, so every page can come directly from
415  * this cache when it is sized to be a multiple of the napi budget.
416  */
417 bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
418 			struct mlx4_en_rx_alloc *frame)
419 {
420 	struct mlx4_en_page_cache *cache = &ring->page_cache;
421 
422 	if (cache->index >= MLX4_EN_CACHE_SIZE)
423 		return false;
424 
425 	cache->buf[cache->index].page = frame->page;
426 	cache->buf[cache->index].dma = frame->dma;
427 	cache->index++;
428 	return true;
429 }
430 
431 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
432 			     struct mlx4_en_rx_ring **pring,
433 			     u32 size, u16 stride)
434 {
435 	struct mlx4_en_dev *mdev = priv->mdev;
436 	struct mlx4_en_rx_ring *ring = *pring;
437 	struct bpf_prog *old_prog;
438 
439 	old_prog = rcu_dereference_protected(
440 					ring->xdp_prog,
441 					lockdep_is_held(&mdev->state_lock));
442 	if (old_prog)
443 		bpf_prog_put(old_prog);
444 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
445 	vfree(ring->rx_info);
446 	ring->rx_info = NULL;
447 	kfree(ring);
448 	*pring = NULL;
449 }
450 
451 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
452 				struct mlx4_en_rx_ring *ring)
453 {
454 	int i;
455 
456 	for (i = 0; i < ring->page_cache.index; i++) {
457 		dma_unmap_page(priv->ddev, ring->page_cache.buf[i].dma,
458 			       PAGE_SIZE, priv->dma_dir);
459 		put_page(ring->page_cache.buf[i].page);
460 	}
461 	ring->page_cache.index = 0;
462 	mlx4_en_free_rx_buf(priv, ring);
463 	if (ring->stride <= TXBB_SIZE)
464 		ring->buf -= TXBB_SIZE;
465 }
466 
467 
468 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
469 				    struct mlx4_en_rx_alloc *frags,
470 				    struct sk_buff *skb,
471 				    int length)
472 {
473 	const struct mlx4_en_frag_info *frag_info = priv->frag_info;
474 	unsigned int truesize = 0;
475 	int nr, frag_size;
476 	struct page *page;
477 	dma_addr_t dma;
478 	bool release;
479 
480 	/* Collect used fragments while replacing them in the HW descriptors */
481 	for (nr = 0;; frags++) {
482 		frag_size = min_t(int, length, frag_info->frag_size);
483 
484 		page = frags->page;
485 		if (unlikely(!page))
486 			goto fail;
487 
488 		dma = frags->dma;
489 		dma_sync_single_range_for_cpu(priv->ddev, dma, frags->page_offset,
490 					      frag_size, priv->dma_dir);
491 
492 		__skb_fill_page_desc(skb, nr, page, frags->page_offset,
493 				     frag_size);
494 
495 		truesize += frag_info->frag_stride;
496 		if (frag_info->frag_stride == PAGE_SIZE / 2) {
497 			frags->page_offset ^= PAGE_SIZE / 2;
498 			release = page_count(page) != 1 ||
499 				  page_is_pfmemalloc(page) ||
500 				  page_to_nid(page) != numa_mem_id();
501 		} else {
502 			u32 sz_align = ALIGN(frag_size, SMP_CACHE_BYTES);
503 
504 			frags->page_offset += sz_align;
505 			release = frags->page_offset + frag_info->frag_size > PAGE_SIZE;
506 		}
507 		if (release) {
508 			dma_unmap_page(priv->ddev, dma, PAGE_SIZE, priv->dma_dir);
509 			frags->page = NULL;
510 		} else {
511 			page_ref_inc(page);
512 		}
513 
514 		nr++;
515 		length -= frag_size;
516 		if (!length)
517 			break;
518 		frag_info++;
519 	}
520 	skb->truesize += truesize;
521 	return nr;
522 
523 fail:
524 	while (nr > 0) {
525 		nr--;
526 		__skb_frag_unref(skb_shinfo(skb)->frags + nr);
527 	}
528 	return 0;
529 }
530 
531 static void validate_loopback(struct mlx4_en_priv *priv, void *va)
532 {
533 	const unsigned char *data = va + ETH_HLEN;
534 	int i;
535 
536 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++) {
537 		if (data[i] != (unsigned char)i)
538 			return;
539 	}
540 	/* Loopback found */
541 	priv->loopback_ok = 1;
542 }
543 
544 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
545 				      struct mlx4_en_rx_ring *ring)
546 {
547 	u32 missing = ring->actual_size - (ring->prod - ring->cons);
548 
549 	/* Try to batch allocations, but not too much. */
550 	if (missing < 8)
551 		return;
552 	do {
553 		if (mlx4_en_prepare_rx_desc(priv, ring,
554 					    ring->prod & ring->size_mask,
555 					    GFP_ATOMIC | __GFP_COLD |
556 					    __GFP_MEMALLOC))
557 			break;
558 		ring->prod++;
559 	} while (likely(--missing));
560 
561 	mlx4_en_update_rx_prod_db(ring);
562 }
563 
564 /* When hardware doesn't strip the vlan, we need to calculate the checksum
565  * over it and add it to the hardware's checksum calculation
566  */
567 static inline __wsum get_fixed_vlan_csum(__wsum hw_checksum,
568 					 struct vlan_hdr *vlanh)
569 {
570 	return csum_add(hw_checksum, *(__wsum *)vlanh);
571 }
572 
573 /* Although the stack expects checksum which doesn't include the pseudo
574  * header, the HW adds it. To address that, we are subtracting the pseudo
575  * header checksum from the checksum value provided by the HW.
576  */
577 static void get_fixed_ipv4_csum(__wsum hw_checksum, struct sk_buff *skb,
578 				struct iphdr *iph)
579 {
580 	__u16 length_for_csum = 0;
581 	__wsum csum_pseudo_header = 0;
582 
583 	length_for_csum = (be16_to_cpu(iph->tot_len) - (iph->ihl << 2));
584 	csum_pseudo_header = csum_tcpudp_nofold(iph->saddr, iph->daddr,
585 						length_for_csum, iph->protocol, 0);
586 	skb->csum = csum_sub(hw_checksum, csum_pseudo_header);
587 }
588 
589 #if IS_ENABLED(CONFIG_IPV6)
590 /* In IPv6 packets, besides subtracting the pseudo header checksum,
591  * we also compute/add the IP header checksum which
592  * is not added by the HW.
593  */
594 static int get_fixed_ipv6_csum(__wsum hw_checksum, struct sk_buff *skb,
595 			       struct ipv6hdr *ipv6h)
596 {
597 	__wsum csum_pseudo_hdr = 0;
598 
599 	if (unlikely(ipv6h->nexthdr == IPPROTO_FRAGMENT ||
600 		     ipv6h->nexthdr == IPPROTO_HOPOPTS))
601 		return -1;
602 	hw_checksum = csum_add(hw_checksum, (__force __wsum)htons(ipv6h->nexthdr));
603 
604 	csum_pseudo_hdr = csum_partial(&ipv6h->saddr,
605 				       sizeof(ipv6h->saddr) + sizeof(ipv6h->daddr), 0);
606 	csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ipv6h->payload_len);
607 	csum_pseudo_hdr = csum_add(csum_pseudo_hdr, (__force __wsum)ntohs(ipv6h->nexthdr));
608 
609 	skb->csum = csum_sub(hw_checksum, csum_pseudo_hdr);
610 	skb->csum = csum_add(skb->csum, csum_partial(ipv6h, sizeof(struct ipv6hdr), 0));
611 	return 0;
612 }
613 #endif
614 static int check_csum(struct mlx4_cqe *cqe, struct sk_buff *skb, void *va,
615 		      netdev_features_t dev_features)
616 {
617 	__wsum hw_checksum = 0;
618 
619 	void *hdr = (u8 *)va + sizeof(struct ethhdr);
620 
621 	hw_checksum = csum_unfold((__force __sum16)cqe->checksum);
622 
623 	if (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK) &&
624 	    !(dev_features & NETIF_F_HW_VLAN_CTAG_RX)) {
625 		hw_checksum = get_fixed_vlan_csum(hw_checksum, hdr);
626 		hdr += sizeof(struct vlan_hdr);
627 	}
628 
629 	if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4))
630 		get_fixed_ipv4_csum(hw_checksum, skb, hdr);
631 #if IS_ENABLED(CONFIG_IPV6)
632 	else if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV6))
633 		if (unlikely(get_fixed_ipv6_csum(hw_checksum, skb, hdr)))
634 			return -1;
635 #endif
636 	return 0;
637 }
638 
639 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
640 {
641 	struct mlx4_en_priv *priv = netdev_priv(dev);
642 	int factor = priv->cqe_factor;
643 	struct mlx4_en_rx_ring *ring;
644 	struct bpf_prog *xdp_prog;
645 	int cq_ring = cq->ring;
646 	bool doorbell_pending;
647 	struct mlx4_cqe *cqe;
648 	int polled = 0;
649 	int index;
650 
651 	if (unlikely(!priv->port_up))
652 		return 0;
653 
654 	if (unlikely(budget <= 0))
655 		return polled;
656 
657 	ring = priv->rx_ring[cq_ring];
658 
659 	/* Protect accesses to: ring->xdp_prog, priv->mac_hash list */
660 	rcu_read_lock();
661 	xdp_prog = rcu_dereference(ring->xdp_prog);
662 	doorbell_pending = 0;
663 
664 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
665 	 * descriptor offset can be deduced from the CQE index instead of
666 	 * reading 'cqe->index' */
667 	index = cq->mcq.cons_index & ring->size_mask;
668 	cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
669 
670 	/* Process all completed CQEs */
671 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
672 		    cq->mcq.cons_index & cq->size)) {
673 		struct mlx4_en_rx_alloc *frags;
674 		enum pkt_hash_types hash_type;
675 		struct sk_buff *skb;
676 		unsigned int length;
677 		int ip_summed;
678 		void *va;
679 		int nr;
680 
681 		frags = ring->rx_info + (index << priv->log_rx_info);
682 		va = page_address(frags[0].page) + frags[0].page_offset;
683 		prefetchw(va);
684 		/*
685 		 * make sure we read the CQE after we read the ownership bit
686 		 */
687 		dma_rmb();
688 
689 		/* Drop packet on bad receive or bad checksum */
690 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
691 						MLX4_CQE_OPCODE_ERROR)) {
692 			en_err(priv, "CQE completed in error - vendor syndrom:%d syndrom:%d\n",
693 			       ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
694 			       ((struct mlx4_err_cqe *)cqe)->syndrome);
695 			goto next;
696 		}
697 		if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
698 			en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
699 			goto next;
700 		}
701 
702 		/* Check if we need to drop the packet if SRIOV is not enabled
703 		 * and not performing the selftest or flb disabled
704 		 */
705 		if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
706 			const struct ethhdr *ethh = va;
707 			dma_addr_t dma;
708 			/* Get pointer to first fragment since we haven't
709 			 * skb yet and cast it to ethhdr struct
710 			 */
711 			dma = frags[0].dma + frags[0].page_offset;
712 			dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
713 						DMA_FROM_DEVICE);
714 
715 			if (is_multicast_ether_addr(ethh->h_dest)) {
716 				struct mlx4_mac_entry *entry;
717 				struct hlist_head *bucket;
718 				unsigned int mac_hash;
719 
720 				/* Drop the packet, since HW loopback-ed it */
721 				mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
722 				bucket = &priv->mac_hash[mac_hash];
723 				hlist_for_each_entry_rcu(entry, bucket, hlist) {
724 					if (ether_addr_equal_64bits(entry->mac,
725 								    ethh->h_source))
726 						goto next;
727 				}
728 			}
729 		}
730 
731 		if (unlikely(priv->validate_loopback)) {
732 			validate_loopback(priv, va);
733 			goto next;
734 		}
735 
736 		/*
737 		 * Packet is OK - process it.
738 		 */
739 		length = be32_to_cpu(cqe->byte_cnt);
740 		length -= ring->fcs_del;
741 
742 		/* A bpf program gets first chance to drop the packet. It may
743 		 * read bytes but not past the end of the frag.
744 		 */
745 		if (xdp_prog) {
746 			struct xdp_buff xdp;
747 			dma_addr_t dma;
748 			void *orig_data;
749 			u32 act;
750 
751 			dma = frags[0].dma + frags[0].page_offset;
752 			dma_sync_single_for_cpu(priv->ddev, dma,
753 						priv->frag_info[0].frag_size,
754 						DMA_FROM_DEVICE);
755 
756 			xdp.data_hard_start = va - frags[0].page_offset;
757 			xdp.data = va;
758 			xdp.data_end = xdp.data + length;
759 			orig_data = xdp.data;
760 
761 			act = bpf_prog_run_xdp(xdp_prog, &xdp);
762 
763 			if (xdp.data != orig_data) {
764 				length = xdp.data_end - xdp.data;
765 				frags[0].page_offset = xdp.data -
766 					xdp.data_hard_start;
767 				va = xdp.data;
768 			}
769 
770 			switch (act) {
771 			case XDP_PASS:
772 				break;
773 			case XDP_TX:
774 				if (likely(!mlx4_en_xmit_frame(ring, frags, dev,
775 							length, cq_ring,
776 							&doorbell_pending))) {
777 					frags[0].page = NULL;
778 					goto next;
779 				}
780 				trace_xdp_exception(dev, xdp_prog, act);
781 				goto xdp_drop_no_cnt; /* Drop on xmit failure */
782 			default:
783 				bpf_warn_invalid_xdp_action(act);
784 			case XDP_ABORTED:
785 				trace_xdp_exception(dev, xdp_prog, act);
786 			case XDP_DROP:
787 				ring->xdp_drop++;
788 xdp_drop_no_cnt:
789 				goto next;
790 			}
791 		}
792 
793 		ring->bytes += length;
794 		ring->packets++;
795 
796 		skb = napi_get_frags(&cq->napi);
797 		if (unlikely(!skb))
798 			goto next;
799 
800 		if (unlikely(ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL)) {
801 			u64 timestamp = mlx4_en_get_cqe_ts(cqe);
802 
803 			mlx4_en_fill_hwtstamps(priv->mdev, skb_hwtstamps(skb),
804 					       timestamp);
805 		}
806 		skb_record_rx_queue(skb, cq_ring);
807 
808 		if (likely(dev->features & NETIF_F_RXCSUM)) {
809 			if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
810 						      MLX4_CQE_STATUS_UDP)) {
811 				if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
812 				    cqe->checksum == cpu_to_be16(0xffff)) {
813 					bool l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
814 						(cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
815 
816 					ip_summed = CHECKSUM_UNNECESSARY;
817 					hash_type = PKT_HASH_TYPE_L4;
818 					if (l2_tunnel)
819 						skb->csum_level = 1;
820 					ring->csum_ok++;
821 				} else {
822 					goto csum_none;
823 				}
824 			} else {
825 				if (priv->flags & MLX4_EN_FLAG_RX_CSUM_NON_TCP_UDP &&
826 				    (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 |
827 							       MLX4_CQE_STATUS_IPV6))) {
828 					if (check_csum(cqe, skb, va, dev->features)) {
829 						goto csum_none;
830 					} else {
831 						ip_summed = CHECKSUM_COMPLETE;
832 						hash_type = PKT_HASH_TYPE_L3;
833 						ring->csum_complete++;
834 					}
835 				} else {
836 					goto csum_none;
837 				}
838 			}
839 		} else {
840 csum_none:
841 			ip_summed = CHECKSUM_NONE;
842 			hash_type = PKT_HASH_TYPE_L3;
843 			ring->csum_none++;
844 		}
845 		skb->ip_summed = ip_summed;
846 		if (dev->features & NETIF_F_RXHASH)
847 			skb_set_hash(skb,
848 				     be32_to_cpu(cqe->immed_rss_invalid),
849 				     hash_type);
850 
851 		if ((cqe->vlan_my_qpn &
852 		     cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK)) &&
853 		    (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
854 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
855 					       be16_to_cpu(cqe->sl_vid));
856 		else if ((cqe->vlan_my_qpn &
857 			  cpu_to_be32(MLX4_CQE_SVLAN_PRESENT_MASK)) &&
858 			 (dev->features & NETIF_F_HW_VLAN_STAG_RX))
859 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021AD),
860 					       be16_to_cpu(cqe->sl_vid));
861 
862 		nr = mlx4_en_complete_rx_desc(priv, frags, skb, length);
863 		if (likely(nr)) {
864 			skb_shinfo(skb)->nr_frags = nr;
865 			skb->len = length;
866 			skb->data_len = length;
867 			napi_gro_frags(&cq->napi);
868 		} else {
869 			skb->vlan_tci = 0;
870 			skb_clear_hash(skb);
871 		}
872 next:
873 		++cq->mcq.cons_index;
874 		index = (cq->mcq.cons_index) & ring->size_mask;
875 		cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
876 		if (unlikely(++polled == budget))
877 			break;
878 	}
879 
880 	rcu_read_unlock();
881 
882 	if (likely(polled)) {
883 		if (doorbell_pending) {
884 			priv->tx_cq[TX_XDP][cq_ring]->xdp_busy = true;
885 			mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq_ring]);
886 		}
887 
888 		mlx4_cq_set_ci(&cq->mcq);
889 		wmb(); /* ensure HW sees CQ consumer before we post new buffers */
890 		ring->cons = cq->mcq.cons_index;
891 	}
892 	AVG_PERF_COUNTER(priv->pstats.rx_coal_avg, polled);
893 
894 	mlx4_en_refill_rx_buffers(priv, ring);
895 
896 	return polled;
897 }
898 
899 
900 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
901 {
902 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
903 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
904 
905 	if (likely(priv->port_up))
906 		napi_schedule_irqoff(&cq->napi);
907 	else
908 		mlx4_en_arm_cq(priv, cq);
909 }
910 
911 /* Rx CQ polling - called by NAPI */
912 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
913 {
914 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
915 	struct net_device *dev = cq->dev;
916 	struct mlx4_en_priv *priv = netdev_priv(dev);
917 	struct mlx4_en_cq *xdp_tx_cq = NULL;
918 	bool clean_complete = true;
919 	int done;
920 
921 	if (priv->tx_ring_num[TX_XDP]) {
922 		xdp_tx_cq = priv->tx_cq[TX_XDP][cq->ring];
923 		if (xdp_tx_cq->xdp_busy) {
924 			clean_complete = mlx4_en_process_tx_cq(dev, xdp_tx_cq,
925 							       budget);
926 			xdp_tx_cq->xdp_busy = !clean_complete;
927 		}
928 	}
929 
930 	done = mlx4_en_process_rx_cq(dev, cq, budget);
931 
932 	/* If we used up all the quota - we're probably not done yet... */
933 	if (done == budget || !clean_complete) {
934 		const struct cpumask *aff;
935 		struct irq_data *idata;
936 		int cpu_curr;
937 
938 		/* in case we got here because of !clean_complete */
939 		done = budget;
940 
941 		INC_PERF_COUNTER(priv->pstats.napi_quota);
942 
943 		cpu_curr = smp_processor_id();
944 		idata = irq_desc_get_irq_data(cq->irq_desc);
945 		aff = irq_data_get_affinity_mask(idata);
946 
947 		if (likely(cpumask_test_cpu(cpu_curr, aff)))
948 			return budget;
949 
950 		/* Current cpu is not according to smp_irq_affinity -
951 		 * probably affinity changed. Need to stop this NAPI
952 		 * poll, and restart it on the right CPU.
953 		 * Try to avoid returning a too small value (like 0),
954 		 * to not fool net_rx_action() and its netdev_budget
955 		 */
956 		if (done)
957 			done--;
958 	}
959 	/* Done for now */
960 	if (likely(napi_complete_done(napi, done)))
961 		mlx4_en_arm_cq(priv, cq);
962 	return done;
963 }
964 
965 void mlx4_en_calc_rx_buf(struct net_device *dev)
966 {
967 	struct mlx4_en_priv *priv = netdev_priv(dev);
968 	int eff_mtu = MLX4_EN_EFF_MTU(dev->mtu);
969 	int i = 0;
970 
971 	/* bpf requires buffers to be set up as 1 packet per page.
972 	 * This only works when num_frags == 1.
973 	 */
974 	if (priv->tx_ring_num[TX_XDP]) {
975 		priv->frag_info[0].frag_size = eff_mtu;
976 		/* This will gain efficient xdp frame recycling at the
977 		 * expense of more costly truesize accounting
978 		 */
979 		priv->frag_info[0].frag_stride = PAGE_SIZE;
980 		priv->dma_dir = PCI_DMA_BIDIRECTIONAL;
981 		priv->rx_headroom = XDP_PACKET_HEADROOM;
982 		i = 1;
983 	} else {
984 		int frag_size_max = 2048, buf_size = 0;
985 
986 		/* should not happen, right ? */
987 		if (eff_mtu > PAGE_SIZE + (MLX4_EN_MAX_RX_FRAGS - 1) * 2048)
988 			frag_size_max = PAGE_SIZE;
989 
990 		while (buf_size < eff_mtu) {
991 			int frag_stride, frag_size = eff_mtu - buf_size;
992 			int pad, nb;
993 
994 			if (i < MLX4_EN_MAX_RX_FRAGS - 1)
995 				frag_size = min(frag_size, frag_size_max);
996 
997 			priv->frag_info[i].frag_size = frag_size;
998 			frag_stride = ALIGN(frag_size, SMP_CACHE_BYTES);
999 			/* We can only pack 2 1536-bytes frames in on 4K page
1000 			 * Therefore, each frame would consume more bytes (truesize)
1001 			 */
1002 			nb = PAGE_SIZE / frag_stride;
1003 			pad = (PAGE_SIZE - nb * frag_stride) / nb;
1004 			pad &= ~(SMP_CACHE_BYTES - 1);
1005 			priv->frag_info[i].frag_stride = frag_stride + pad;
1006 
1007 			buf_size += frag_size;
1008 			i++;
1009 		}
1010 		priv->dma_dir = PCI_DMA_FROMDEVICE;
1011 		priv->rx_headroom = 0;
1012 	}
1013 
1014 	priv->num_frags = i;
1015 	priv->rx_skb_size = eff_mtu;
1016 	priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
1017 
1018 	en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d num_frags:%d):\n",
1019 	       eff_mtu, priv->num_frags);
1020 	for (i = 0; i < priv->num_frags; i++) {
1021 		en_dbg(DRV,
1022 		       priv,
1023 		       "  frag:%d - size:%d stride:%d\n",
1024 		       i,
1025 		       priv->frag_info[i].frag_size,
1026 		       priv->frag_info[i].frag_stride);
1027 	}
1028 }
1029 
1030 /* RSS related functions */
1031 
1032 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
1033 				 struct mlx4_en_rx_ring *ring,
1034 				 enum mlx4_qp_state *state,
1035 				 struct mlx4_qp *qp)
1036 {
1037 	struct mlx4_en_dev *mdev = priv->mdev;
1038 	struct mlx4_qp_context *context;
1039 	int err = 0;
1040 
1041 	context = kmalloc(sizeof(*context), GFP_KERNEL);
1042 	if (!context)
1043 		return -ENOMEM;
1044 
1045 	err = mlx4_qp_alloc(mdev->dev, qpn, qp, GFP_KERNEL);
1046 	if (err) {
1047 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
1048 		goto out;
1049 	}
1050 	qp->event = mlx4_en_sqp_event;
1051 
1052 	memset(context, 0, sizeof *context);
1053 	mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
1054 				qpn, ring->cqn, -1, context);
1055 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
1056 
1057 	/* Cancel FCS removal if FW allows */
1058 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
1059 		context->param3 |= cpu_to_be32(1 << 29);
1060 		if (priv->dev->features & NETIF_F_RXFCS)
1061 			ring->fcs_del = 0;
1062 		else
1063 			ring->fcs_del = ETH_FCS_LEN;
1064 	} else
1065 		ring->fcs_del = 0;
1066 
1067 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
1068 	if (err) {
1069 		mlx4_qp_remove(mdev->dev, qp);
1070 		mlx4_qp_free(mdev->dev, qp);
1071 	}
1072 	mlx4_en_update_rx_prod_db(ring);
1073 out:
1074 	kfree(context);
1075 	return err;
1076 }
1077 
1078 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
1079 {
1080 	int err;
1081 	u32 qpn;
1082 
1083 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn,
1084 				    MLX4_RESERVE_A0_QP);
1085 	if (err) {
1086 		en_err(priv, "Failed reserving drop qpn\n");
1087 		return err;
1088 	}
1089 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp, GFP_KERNEL);
1090 	if (err) {
1091 		en_err(priv, "Failed allocating drop qp\n");
1092 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1093 		return err;
1094 	}
1095 
1096 	return 0;
1097 }
1098 
1099 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
1100 {
1101 	u32 qpn;
1102 
1103 	qpn = priv->drop_qp.qpn;
1104 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1105 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1106 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1107 }
1108 
1109 /* Allocate rx qp's and configure them according to rss map */
1110 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1111 {
1112 	struct mlx4_en_dev *mdev = priv->mdev;
1113 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1114 	struct mlx4_qp_context context;
1115 	struct mlx4_rss_context *rss_context;
1116 	int rss_rings;
1117 	void *ptr;
1118 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1119 			MLX4_RSS_TCP_IPV6);
1120 	int i, qpn;
1121 	int err = 0;
1122 	int good_qps = 0;
1123 	u8 flags;
1124 
1125 	en_dbg(DRV, priv, "Configuring rss steering\n");
1126 
1127 	flags = priv->rx_ring_num == 1 ? MLX4_RESERVE_A0_QP : 0;
1128 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1129 				    priv->rx_ring_num,
1130 				    &rss_map->base_qpn, flags);
1131 	if (err) {
1132 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1133 		return err;
1134 	}
1135 
1136 	for (i = 0; i < priv->rx_ring_num; i++) {
1137 		qpn = rss_map->base_qpn + i;
1138 		err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1139 					    &rss_map->state[i],
1140 					    &rss_map->qps[i]);
1141 		if (err)
1142 			goto rss_err;
1143 
1144 		++good_qps;
1145 	}
1146 
1147 	if (priv->rx_ring_num == 1) {
1148 		rss_map->indir_qp = &rss_map->qps[0];
1149 		priv->base_qpn = rss_map->indir_qp->qpn;
1150 		en_info(priv, "Optimized Non-RSS steering\n");
1151 		return 0;
1152 	}
1153 
1154 	rss_map->indir_qp = kzalloc(sizeof(*rss_map->indir_qp), GFP_KERNEL);
1155 	if (!rss_map->indir_qp) {
1156 		err = -ENOMEM;
1157 		goto rss_err;
1158 	}
1159 
1160 	/* Configure RSS indirection qp */
1161 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, rss_map->indir_qp,
1162 			    GFP_KERNEL);
1163 	if (err) {
1164 		en_err(priv, "Failed to allocate RSS indirection QP\n");
1165 		goto rss_err;
1166 	}
1167 
1168 	rss_map->indir_qp->event = mlx4_en_sqp_event;
1169 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1170 				priv->rx_ring[0]->cqn, -1, &context);
1171 
1172 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1173 		rss_rings = priv->rx_ring_num;
1174 	else
1175 		rss_rings = priv->prof->rss_rings;
1176 
1177 	ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1178 					+ MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1179 	rss_context = ptr;
1180 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1181 					    (rss_map->base_qpn));
1182 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1183 	if (priv->mdev->profile.udp_rss) {
1184 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1185 		rss_context->base_qpn_udp = rss_context->default_qpn;
1186 	}
1187 
1188 	if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1189 		en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1190 		rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1191 	}
1192 
1193 	rss_context->flags = rss_mask;
1194 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1195 	if (priv->rss_hash_fn == ETH_RSS_HASH_XOR) {
1196 		rss_context->hash_fn = MLX4_RSS_HASH_XOR;
1197 	} else if (priv->rss_hash_fn == ETH_RSS_HASH_TOP) {
1198 		rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1199 		memcpy(rss_context->rss_key, priv->rss_key,
1200 		       MLX4_EN_RSS_KEY_SIZE);
1201 	} else {
1202 		en_err(priv, "Unknown RSS hash function requested\n");
1203 		err = -EINVAL;
1204 		goto indir_err;
1205 	}
1206 
1207 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1208 			       rss_map->indir_qp, &rss_map->indir_state);
1209 	if (err)
1210 		goto indir_err;
1211 
1212 	return 0;
1213 
1214 indir_err:
1215 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1216 		       MLX4_QP_STATE_RST, NULL, 0, 0, rss_map->indir_qp);
1217 	mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
1218 	mlx4_qp_free(mdev->dev, rss_map->indir_qp);
1219 	kfree(rss_map->indir_qp);
1220 	rss_map->indir_qp = NULL;
1221 rss_err:
1222 	for (i = 0; i < good_qps; i++) {
1223 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1224 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1225 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1226 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1227 	}
1228 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1229 	return err;
1230 }
1231 
1232 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1233 {
1234 	struct mlx4_en_dev *mdev = priv->mdev;
1235 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1236 	int i;
1237 
1238 	if (priv->rx_ring_num > 1) {
1239 		mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1240 			       MLX4_QP_STATE_RST, NULL, 0, 0,
1241 			       rss_map->indir_qp);
1242 		mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
1243 		mlx4_qp_free(mdev->dev, rss_map->indir_qp);
1244 		kfree(rss_map->indir_qp);
1245 		rss_map->indir_qp = NULL;
1246 	}
1247 
1248 	for (i = 0; i < priv->rx_ring_num; i++) {
1249 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1250 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1251 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1252 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1253 	}
1254 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1255 }
1256