xref: /linux/drivers/net/ethernet/mellanox/mlx4/en_rx.c (revision 55d0969c451159cff86949b38c39171cab962069)
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 <linux/bpf.h>
35 #include <linux/bpf_trace.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 #include <linux/skbuff_ref.h>
46 
47 #include <net/ip.h>
48 #if IS_ENABLED(CONFIG_IPV6)
49 #include <net/ip6_checksum.h>
50 #endif
51 
52 #include "mlx4_en.h"
53 
54 static int mlx4_alloc_page(struct mlx4_en_priv *priv,
55 			   struct mlx4_en_rx_alloc *frag,
56 			   gfp_t gfp)
57 {
58 	struct page *page;
59 	dma_addr_t dma;
60 
61 	page = alloc_page(gfp);
62 	if (unlikely(!page))
63 		return -ENOMEM;
64 	dma = dma_map_page(priv->ddev, page, 0, PAGE_SIZE, priv->dma_dir);
65 	if (unlikely(dma_mapping_error(priv->ddev, dma))) {
66 		__free_page(page);
67 		return -ENOMEM;
68 	}
69 	frag->page = page;
70 	frag->dma = dma;
71 	frag->page_offset = priv->rx_headroom;
72 	return 0;
73 }
74 
75 static int mlx4_en_alloc_frags(struct mlx4_en_priv *priv,
76 			       struct mlx4_en_rx_ring *ring,
77 			       struct mlx4_en_rx_desc *rx_desc,
78 			       struct mlx4_en_rx_alloc *frags,
79 			       gfp_t gfp)
80 {
81 	int i;
82 
83 	for (i = 0; i < priv->num_frags; i++, frags++) {
84 		if (!frags->page) {
85 			if (mlx4_alloc_page(priv, frags, gfp)) {
86 				ring->alloc_fail++;
87 				return -ENOMEM;
88 			}
89 			ring->rx_alloc_pages++;
90 		}
91 		rx_desc->data[i].addr = cpu_to_be64(frags->dma +
92 						    frags->page_offset);
93 	}
94 	return 0;
95 }
96 
97 static void mlx4_en_free_frag(const struct mlx4_en_priv *priv,
98 			      struct mlx4_en_rx_alloc *frag)
99 {
100 	if (frag->page) {
101 		dma_unmap_page(priv->ddev, frag->dma,
102 			       PAGE_SIZE, priv->dma_dir);
103 		__free_page(frag->page);
104 	}
105 	/* We need to clear all fields, otherwise a change of priv->log_rx_info
106 	 * could lead to see garbage later in frag->page.
107 	 */
108 	memset(frag, 0, sizeof(*frag));
109 }
110 
111 static void mlx4_en_init_rx_desc(const struct mlx4_en_priv *priv,
112 				 struct mlx4_en_rx_ring *ring, int index)
113 {
114 	struct mlx4_en_rx_desc *rx_desc = ring->buf + ring->stride * index;
115 	int possible_frags;
116 	int i;
117 
118 	/* Set size and memtype fields */
119 	for (i = 0; i < priv->num_frags; i++) {
120 		rx_desc->data[i].byte_count =
121 			cpu_to_be32(priv->frag_info[i].frag_size);
122 		rx_desc->data[i].lkey = cpu_to_be32(priv->mdev->mr.key);
123 	}
124 
125 	/* If the number of used fragments does not fill up the ring stride,
126 	 * remaining (unused) fragments must be padded with null address/size
127 	 * and a special memory key */
128 	possible_frags = (ring->stride - sizeof(struct mlx4_en_rx_desc)) / DS_SIZE;
129 	for (i = priv->num_frags; i < possible_frags; i++) {
130 		rx_desc->data[i].byte_count = 0;
131 		rx_desc->data[i].lkey = cpu_to_be32(MLX4_EN_MEMTYPE_PAD);
132 		rx_desc->data[i].addr = 0;
133 	}
134 }
135 
136 static int mlx4_en_prepare_rx_desc(struct mlx4_en_priv *priv,
137 				   struct mlx4_en_rx_ring *ring, int index,
138 				   gfp_t gfp)
139 {
140 	struct mlx4_en_rx_desc *rx_desc = ring->buf +
141 		(index << ring->log_stride);
142 	struct mlx4_en_rx_alloc *frags = ring->rx_info +
143 					(index << priv->log_rx_info);
144 	if (likely(ring->page_cache.index > 0)) {
145 		/* XDP uses a single page per frame */
146 		if (!frags->page) {
147 			ring->page_cache.index--;
148 			frags->page = ring->page_cache.buf[ring->page_cache.index].page;
149 			frags->dma  = ring->page_cache.buf[ring->page_cache.index].dma;
150 		}
151 		frags->page_offset = XDP_PACKET_HEADROOM;
152 		rx_desc->data[0].addr = cpu_to_be64(frags->dma +
153 						    XDP_PACKET_HEADROOM);
154 		return 0;
155 	}
156 
157 	return mlx4_en_alloc_frags(priv, ring, rx_desc, frags, gfp);
158 }
159 
160 static bool mlx4_en_is_ring_empty(const struct mlx4_en_rx_ring *ring)
161 {
162 	return ring->prod == ring->cons;
163 }
164 
165 static inline void mlx4_en_update_rx_prod_db(struct mlx4_en_rx_ring *ring)
166 {
167 	*ring->wqres.db.db = cpu_to_be32(ring->prod & 0xffff);
168 }
169 
170 /* slow path */
171 static void mlx4_en_free_rx_desc(const struct mlx4_en_priv *priv,
172 				 struct mlx4_en_rx_ring *ring,
173 				 int index)
174 {
175 	struct mlx4_en_rx_alloc *frags;
176 	int nr;
177 
178 	frags = ring->rx_info + (index << priv->log_rx_info);
179 	for (nr = 0; nr < priv->num_frags; nr++) {
180 		en_dbg(DRV, priv, "Freeing fragment:%d\n", nr);
181 		mlx4_en_free_frag(priv, frags + nr);
182 	}
183 }
184 
185 /* Function not in fast-path */
186 static int mlx4_en_fill_rx_buffers(struct mlx4_en_priv *priv)
187 {
188 	struct mlx4_en_rx_ring *ring;
189 	int ring_ind;
190 	int buf_ind;
191 	int new_size;
192 
193 	for (buf_ind = 0; buf_ind < priv->prof->rx_ring_size; buf_ind++) {
194 		for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
195 			ring = priv->rx_ring[ring_ind];
196 
197 			if (mlx4_en_prepare_rx_desc(priv, ring,
198 						    ring->actual_size,
199 						    GFP_KERNEL)) {
200 				if (ring->actual_size < MLX4_EN_MIN_RX_SIZE) {
201 					en_err(priv, "Failed to allocate enough rx buffers\n");
202 					return -ENOMEM;
203 				} else {
204 					new_size = rounddown_pow_of_two(ring->actual_size);
205 					en_warn(priv, "Only %d buffers allocated reducing ring size to %d\n",
206 						ring->actual_size, new_size);
207 					goto reduce_rings;
208 				}
209 			}
210 			ring->actual_size++;
211 			ring->prod++;
212 		}
213 	}
214 	return 0;
215 
216 reduce_rings:
217 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
218 		ring = priv->rx_ring[ring_ind];
219 		while (ring->actual_size > new_size) {
220 			ring->actual_size--;
221 			ring->prod--;
222 			mlx4_en_free_rx_desc(priv, ring, ring->actual_size);
223 		}
224 	}
225 
226 	return 0;
227 }
228 
229 static void mlx4_en_free_rx_buf(struct mlx4_en_priv *priv,
230 				struct mlx4_en_rx_ring *ring)
231 {
232 	int index;
233 
234 	en_dbg(DRV, priv, "Freeing Rx buf - cons:%d prod:%d\n",
235 	       ring->cons, ring->prod);
236 
237 	/* Unmap and free Rx buffers */
238 	for (index = 0; index < ring->size; index++) {
239 		en_dbg(DRV, priv, "Processing descriptor:%d\n", index);
240 		mlx4_en_free_rx_desc(priv, ring, index);
241 	}
242 	ring->cons = 0;
243 	ring->prod = 0;
244 }
245 
246 void mlx4_en_set_num_rx_rings(struct mlx4_en_dev *mdev)
247 {
248 	int i;
249 	int num_of_eqs;
250 	int num_rx_rings;
251 	struct mlx4_dev *dev = mdev->dev;
252 
253 	mlx4_foreach_port(i, dev, MLX4_PORT_TYPE_ETH) {
254 		num_of_eqs = max_t(int, MIN_RX_RINGS,
255 				   min_t(int,
256 					 mlx4_get_eqs_per_port(mdev->dev, i),
257 					 DEF_RX_RINGS));
258 
259 		num_rx_rings = mlx4_low_memory_profile() ? MIN_RX_RINGS :
260 			min_t(int, num_of_eqs, num_online_cpus());
261 		mdev->profile.prof[i].rx_ring_num =
262 			rounddown_pow_of_two(num_rx_rings);
263 	}
264 }
265 
266 int mlx4_en_create_rx_ring(struct mlx4_en_priv *priv,
267 			   struct mlx4_en_rx_ring **pring,
268 			   u32 size, u16 stride, int node, int queue_index)
269 {
270 	struct mlx4_en_dev *mdev = priv->mdev;
271 	struct mlx4_en_rx_ring *ring;
272 	int err = -ENOMEM;
273 	int tmp;
274 
275 	ring = kzalloc_node(sizeof(*ring), GFP_KERNEL, node);
276 	if (!ring) {
277 		en_err(priv, "Failed to allocate RX ring structure\n");
278 		return -ENOMEM;
279 	}
280 
281 	ring->prod = 0;
282 	ring->cons = 0;
283 	ring->size = size;
284 	ring->size_mask = size - 1;
285 	ring->stride = stride;
286 	ring->log_stride = ffs(ring->stride) - 1;
287 	ring->buf_size = ring->size * ring->stride + TXBB_SIZE;
288 
289 	if (xdp_rxq_info_reg(&ring->xdp_rxq, priv->dev, queue_index, 0) < 0)
290 		goto err_ring;
291 
292 	tmp = size * roundup_pow_of_two(MLX4_EN_MAX_RX_FRAGS *
293 					sizeof(struct mlx4_en_rx_alloc));
294 	ring->rx_info = kvzalloc_node(tmp, GFP_KERNEL, node);
295 	if (!ring->rx_info) {
296 		err = -ENOMEM;
297 		goto err_xdp_info;
298 	}
299 
300 	en_dbg(DRV, priv, "Allocated rx_info ring at addr:%p size:%d\n",
301 		 ring->rx_info, tmp);
302 
303 	/* Allocate HW buffers on provided NUMA node */
304 	set_dev_node(&mdev->dev->persist->pdev->dev, node);
305 	err = mlx4_alloc_hwq_res(mdev->dev, &ring->wqres, ring->buf_size);
306 	set_dev_node(&mdev->dev->persist->pdev->dev, mdev->dev->numa_node);
307 	if (err)
308 		goto err_info;
309 
310 	ring->buf = ring->wqres.buf.direct.buf;
311 
312 	ring->hwtstamp_rx_filter = priv->hwtstamp_config.rx_filter;
313 
314 	*pring = ring;
315 	return 0;
316 
317 err_info:
318 	kvfree(ring->rx_info);
319 	ring->rx_info = NULL;
320 err_xdp_info:
321 	xdp_rxq_info_unreg(&ring->xdp_rxq);
322 err_ring:
323 	kfree(ring);
324 	*pring = NULL;
325 
326 	return err;
327 }
328 
329 int mlx4_en_activate_rx_rings(struct mlx4_en_priv *priv)
330 {
331 	struct mlx4_en_rx_ring *ring;
332 	int i;
333 	int ring_ind;
334 	int err;
335 	int stride = roundup_pow_of_two(sizeof(struct mlx4_en_rx_desc) +
336 					DS_SIZE * priv->num_frags);
337 
338 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
339 		ring = priv->rx_ring[ring_ind];
340 
341 		ring->prod = 0;
342 		ring->cons = 0;
343 		ring->actual_size = 0;
344 		ring->cqn = priv->rx_cq[ring_ind]->mcq.cqn;
345 
346 		ring->stride = stride;
347 		if (ring->stride <= TXBB_SIZE) {
348 			/* Stamp first unused send wqe */
349 			__be32 *ptr = (__be32 *)ring->buf;
350 			__be32 stamp = cpu_to_be32(1 << STAMP_SHIFT);
351 			*ptr = stamp;
352 			/* Move pointer to start of rx section */
353 			ring->buf += TXBB_SIZE;
354 		}
355 
356 		ring->log_stride = ffs(ring->stride) - 1;
357 		ring->buf_size = ring->size * ring->stride;
358 
359 		memset(ring->buf, 0, ring->buf_size);
360 		mlx4_en_update_rx_prod_db(ring);
361 
362 		/* Initialize all descriptors */
363 		for (i = 0; i < ring->size; i++)
364 			mlx4_en_init_rx_desc(priv, ring, i);
365 	}
366 	err = mlx4_en_fill_rx_buffers(priv);
367 	if (err)
368 		goto err_buffers;
369 
370 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++) {
371 		ring = priv->rx_ring[ring_ind];
372 
373 		ring->size_mask = ring->actual_size - 1;
374 		mlx4_en_update_rx_prod_db(ring);
375 	}
376 
377 	return 0;
378 
379 err_buffers:
380 	for (ring_ind = 0; ring_ind < priv->rx_ring_num; ring_ind++)
381 		mlx4_en_free_rx_buf(priv, priv->rx_ring[ring_ind]);
382 
383 	ring_ind = priv->rx_ring_num - 1;
384 	while (ring_ind >= 0) {
385 		if (priv->rx_ring[ring_ind]->stride <= TXBB_SIZE)
386 			priv->rx_ring[ring_ind]->buf -= TXBB_SIZE;
387 		ring_ind--;
388 	}
389 	return err;
390 }
391 
392 /* We recover from out of memory by scheduling our napi poll
393  * function (mlx4_en_process_cq), which tries to allocate
394  * all missing RX buffers (call to mlx4_en_refill_rx_buffers).
395  */
396 void mlx4_en_recover_from_oom(struct mlx4_en_priv *priv)
397 {
398 	int ring;
399 
400 	if (!priv->port_up)
401 		return;
402 
403 	for (ring = 0; ring < priv->rx_ring_num; ring++) {
404 		if (mlx4_en_is_ring_empty(priv->rx_ring[ring])) {
405 			local_bh_disable();
406 			napi_schedule(&priv->rx_cq[ring]->napi);
407 			local_bh_enable();
408 		}
409 	}
410 }
411 
412 /* When the rx ring is running in page-per-packet mode, a released frame can go
413  * directly into a small cache, to avoid unmapping or touching the page
414  * allocator. In bpf prog performance scenarios, buffers are either forwarded
415  * or dropped, never converted to skbs, so every page can come directly from
416  * this cache when it is sized to be a multiple of the napi budget.
417  */
418 bool mlx4_en_rx_recycle(struct mlx4_en_rx_ring *ring,
419 			struct mlx4_en_rx_alloc *frame)
420 {
421 	struct mlx4_en_page_cache *cache = &ring->page_cache;
422 
423 	if (cache->index >= MLX4_EN_CACHE_SIZE)
424 		return false;
425 
426 	cache->buf[cache->index].page = frame->page;
427 	cache->buf[cache->index].dma = frame->dma;
428 	cache->index++;
429 	return true;
430 }
431 
432 void mlx4_en_destroy_rx_ring(struct mlx4_en_priv *priv,
433 			     struct mlx4_en_rx_ring **pring,
434 			     u32 size, u16 stride)
435 {
436 	struct mlx4_en_dev *mdev = priv->mdev;
437 	struct mlx4_en_rx_ring *ring = *pring;
438 	struct bpf_prog *old_prog;
439 
440 	old_prog = rcu_dereference_protected(
441 					ring->xdp_prog,
442 					lockdep_is_held(&mdev->state_lock));
443 	if (old_prog)
444 		bpf_prog_put(old_prog);
445 	xdp_rxq_info_unreg(&ring->xdp_rxq);
446 	mlx4_free_hwq_res(mdev->dev, &ring->wqres, size * stride + TXBB_SIZE);
447 	kvfree(ring->rx_info);
448 	ring->rx_info = NULL;
449 	kfree(ring);
450 	*pring = NULL;
451 }
452 
453 void mlx4_en_deactivate_rx_ring(struct mlx4_en_priv *priv,
454 				struct mlx4_en_rx_ring *ring)
455 {
456 	int i;
457 
458 	for (i = 0; i < ring->page_cache.index; i++) {
459 		dma_unmap_page(priv->ddev, ring->page_cache.buf[i].dma,
460 			       PAGE_SIZE, priv->dma_dir);
461 		put_page(ring->page_cache.buf[i].page);
462 	}
463 	ring->page_cache.index = 0;
464 	mlx4_en_free_rx_buf(priv, ring);
465 	if (ring->stride <= TXBB_SIZE)
466 		ring->buf -= TXBB_SIZE;
467 }
468 
469 
470 static int mlx4_en_complete_rx_desc(struct mlx4_en_priv *priv,
471 				    struct mlx4_en_rx_alloc *frags,
472 				    struct sk_buff *skb,
473 				    int length)
474 {
475 	const struct mlx4_en_frag_info *frag_info = priv->frag_info;
476 	unsigned int truesize = 0;
477 	bool release = true;
478 	int nr, frag_size;
479 	struct page *page;
480 	dma_addr_t dma;
481 
482 	/* Collect used fragments while replacing them in the HW descriptors */
483 	for (nr = 0;; frags++) {
484 		frag_size = min_t(int, length, frag_info->frag_size);
485 
486 		page = frags->page;
487 		if (unlikely(!page))
488 			goto fail;
489 
490 		dma = frags->dma;
491 		dma_sync_single_range_for_cpu(priv->ddev, dma, frags->page_offset,
492 					      frag_size, priv->dma_dir);
493 
494 		__skb_fill_page_desc(skb, nr, page, frags->page_offset,
495 				     frag_size);
496 
497 		truesize += frag_info->frag_stride;
498 		if (frag_info->frag_stride == PAGE_SIZE / 2) {
499 			frags->page_offset ^= PAGE_SIZE / 2;
500 			release = page_count(page) != 1 ||
501 				  page_is_pfmemalloc(page) ||
502 				  page_to_nid(page) != numa_mem_id();
503 		} else if (!priv->rx_headroom) {
504 			/* rx_headroom for non XDP setup is always 0.
505 			 * When XDP is set, the above condition will
506 			 * guarantee page is always released.
507 			 */
508 			u32 sz_align = ALIGN(frag_size, SMP_CACHE_BYTES);
509 
510 			frags->page_offset += sz_align;
511 			release = frags->page_offset + frag_info->frag_size > PAGE_SIZE;
512 		}
513 		if (release) {
514 			dma_unmap_page(priv->ddev, dma, PAGE_SIZE, priv->dma_dir);
515 			frags->page = NULL;
516 		} else {
517 			page_ref_inc(page);
518 		}
519 
520 		nr++;
521 		length -= frag_size;
522 		if (!length)
523 			break;
524 		frag_info++;
525 	}
526 	skb->truesize += truesize;
527 	return nr;
528 
529 fail:
530 	while (nr > 0) {
531 		nr--;
532 		__skb_frag_unref(skb_shinfo(skb)->frags + nr, false);
533 	}
534 	return 0;
535 }
536 
537 static void validate_loopback(struct mlx4_en_priv *priv, void *va)
538 {
539 	const unsigned char *data = va + ETH_HLEN;
540 	int i;
541 
542 	for (i = 0; i < MLX4_LOOPBACK_TEST_PAYLOAD; i++) {
543 		if (data[i] != (unsigned char)i)
544 			return;
545 	}
546 	/* Loopback found */
547 	priv->loopback_ok = 1;
548 }
549 
550 static void mlx4_en_refill_rx_buffers(struct mlx4_en_priv *priv,
551 				      struct mlx4_en_rx_ring *ring)
552 {
553 	u32 missing = ring->actual_size - (ring->prod - ring->cons);
554 
555 	/* Try to batch allocations, but not too much. */
556 	if (missing < 8)
557 		return;
558 	do {
559 		if (mlx4_en_prepare_rx_desc(priv, ring,
560 					    ring->prod & ring->size_mask,
561 					    GFP_ATOMIC | __GFP_MEMALLOC))
562 			break;
563 		ring->prod++;
564 	} while (likely(--missing));
565 
566 	mlx4_en_update_rx_prod_db(ring);
567 }
568 
569 /* When hardware doesn't strip the vlan, we need to calculate the checksum
570  * over it and add it to the hardware's checksum calculation
571  */
572 static inline __wsum get_fixed_vlan_csum(__wsum hw_checksum,
573 					 struct vlan_hdr *vlanh)
574 {
575 	return csum_add(hw_checksum, *(__wsum *)vlanh);
576 }
577 
578 /* Although the stack expects checksum which doesn't include the pseudo
579  * header, the HW adds it. To address that, we are subtracting the pseudo
580  * header checksum from the checksum value provided by the HW.
581  */
582 static int get_fixed_ipv4_csum(__wsum hw_checksum, struct sk_buff *skb,
583 			       struct iphdr *iph)
584 {
585 	__u16 length_for_csum = 0;
586 	__wsum csum_pseudo_header = 0;
587 	__u8 ipproto = iph->protocol;
588 
589 	if (unlikely(ipproto == IPPROTO_SCTP))
590 		return -1;
591 
592 	length_for_csum = (be16_to_cpu(iph->tot_len) - (iph->ihl << 2));
593 	csum_pseudo_header = csum_tcpudp_nofold(iph->saddr, iph->daddr,
594 						length_for_csum, ipproto, 0);
595 	skb->csum = csum_sub(hw_checksum, csum_pseudo_header);
596 	return 0;
597 }
598 
599 #if IS_ENABLED(CONFIG_IPV6)
600 /* In IPv6 packets, hw_checksum lacks 6 bytes from IPv6 header:
601  * 4 first bytes : priority, version, flow_lbl
602  * and 2 additional bytes : nexthdr, hop_limit.
603  */
604 static int get_fixed_ipv6_csum(__wsum hw_checksum, struct sk_buff *skb,
605 			       struct ipv6hdr *ipv6h)
606 {
607 	__u8 nexthdr = ipv6h->nexthdr;
608 	__wsum temp;
609 
610 	if (unlikely(nexthdr == IPPROTO_FRAGMENT ||
611 		     nexthdr == IPPROTO_HOPOPTS ||
612 		     nexthdr == IPPROTO_SCTP))
613 		return -1;
614 
615 	/* priority, version, flow_lbl */
616 	temp = csum_add(hw_checksum, *(__wsum *)ipv6h);
617 	/* nexthdr and hop_limit */
618 	skb->csum = csum_add(temp, (__force __wsum)*(__be16 *)&ipv6h->nexthdr);
619 	return 0;
620 }
621 #endif
622 
623 #define short_frame(size) ((size) <= ETH_ZLEN + ETH_FCS_LEN)
624 
625 /* We reach this function only after checking that any of
626  * the (IPv4 | IPv6) bits are set in cqe->status.
627  */
628 static int check_csum(struct mlx4_cqe *cqe, struct sk_buff *skb, void *va,
629 		      netdev_features_t dev_features)
630 {
631 	__wsum hw_checksum = 0;
632 	void *hdr;
633 
634 	/* CQE csum doesn't cover padding octets in short ethernet
635 	 * frames. And the pad field is appended prior to calculating
636 	 * and appending the FCS field.
637 	 *
638 	 * Detecting these padded frames requires to verify and parse
639 	 * IP headers, so we simply force all those small frames to skip
640 	 * checksum complete.
641 	 */
642 	if (short_frame(skb->len))
643 		return -EINVAL;
644 
645 	hdr = (u8 *)va + sizeof(struct ethhdr);
646 	hw_checksum = csum_unfold((__force __sum16)cqe->checksum);
647 
648 	if (cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK) &&
649 	    !(dev_features & NETIF_F_HW_VLAN_CTAG_RX)) {
650 		hw_checksum = get_fixed_vlan_csum(hw_checksum, hdr);
651 		hdr += sizeof(struct vlan_hdr);
652 	}
653 
654 #if IS_ENABLED(CONFIG_IPV6)
655 	if (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPV6))
656 		return get_fixed_ipv6_csum(hw_checksum, skb, hdr);
657 #endif
658 	return get_fixed_ipv4_csum(hw_checksum, skb, hdr);
659 }
660 
661 #if IS_ENABLED(CONFIG_IPV6)
662 #define MLX4_CQE_STATUS_IP_ANY (MLX4_CQE_STATUS_IPV4 | MLX4_CQE_STATUS_IPV6)
663 #else
664 #define MLX4_CQE_STATUS_IP_ANY (MLX4_CQE_STATUS_IPV4)
665 #endif
666 
667 struct mlx4_en_xdp_buff {
668 	struct xdp_buff xdp;
669 	struct mlx4_cqe *cqe;
670 	struct mlx4_en_dev *mdev;
671 	struct mlx4_en_rx_ring *ring;
672 	struct net_device *dev;
673 };
674 
675 int mlx4_en_xdp_rx_timestamp(const struct xdp_md *ctx, u64 *timestamp)
676 {
677 	struct mlx4_en_xdp_buff *_ctx = (void *)ctx;
678 
679 	if (unlikely(_ctx->ring->hwtstamp_rx_filter != HWTSTAMP_FILTER_ALL))
680 		return -ENODATA;
681 
682 	*timestamp = mlx4_en_get_hwtstamp(_ctx->mdev,
683 					  mlx4_en_get_cqe_ts(_ctx->cqe));
684 	return 0;
685 }
686 
687 int mlx4_en_xdp_rx_hash(const struct xdp_md *ctx, u32 *hash,
688 			enum xdp_rss_hash_type *rss_type)
689 {
690 	struct mlx4_en_xdp_buff *_ctx = (void *)ctx;
691 	struct mlx4_cqe *cqe = _ctx->cqe;
692 	enum xdp_rss_hash_type xht = 0;
693 	__be16 status;
694 
695 	if (unlikely(!(_ctx->dev->features & NETIF_F_RXHASH)))
696 		return -ENODATA;
697 
698 	*hash = be32_to_cpu(cqe->immed_rss_invalid);
699 	status = cqe->status;
700 	if (status & cpu_to_be16(MLX4_CQE_STATUS_TCP))
701 		xht = XDP_RSS_L4_TCP;
702 	if (status & cpu_to_be16(MLX4_CQE_STATUS_UDP))
703 		xht = XDP_RSS_L4_UDP;
704 	if (status & cpu_to_be16(MLX4_CQE_STATUS_IPV4 | MLX4_CQE_STATUS_IPV4F))
705 		xht |= XDP_RSS_L3_IPV4;
706 	if (status & cpu_to_be16(MLX4_CQE_STATUS_IPV6)) {
707 		xht |= XDP_RSS_L3_IPV6;
708 		if (cqe->ipv6_ext_mask)
709 			xht |= XDP_RSS_L3_DYNHDR;
710 	}
711 	*rss_type = xht;
712 
713 	return 0;
714 }
715 
716 int mlx4_en_process_rx_cq(struct net_device *dev, struct mlx4_en_cq *cq, int budget)
717 {
718 	struct mlx4_en_priv *priv = netdev_priv(dev);
719 	struct mlx4_en_xdp_buff mxbuf = {};
720 	int factor = priv->cqe_factor;
721 	struct mlx4_en_rx_ring *ring;
722 	struct bpf_prog *xdp_prog;
723 	int cq_ring = cq->ring;
724 	bool doorbell_pending;
725 	bool xdp_redir_flush;
726 	struct mlx4_cqe *cqe;
727 	int polled = 0;
728 	int index;
729 
730 	if (unlikely(!priv->port_up || budget <= 0))
731 		return 0;
732 
733 	ring = priv->rx_ring[cq_ring];
734 
735 	xdp_prog = rcu_dereference_bh(ring->xdp_prog);
736 	xdp_init_buff(&mxbuf.xdp, priv->frag_info[0].frag_stride, &ring->xdp_rxq);
737 	doorbell_pending = false;
738 	xdp_redir_flush = false;
739 
740 	/* We assume a 1:1 mapping between CQEs and Rx descriptors, so Rx
741 	 * descriptor offset can be deduced from the CQE index instead of
742 	 * reading 'cqe->index' */
743 	index = cq->mcq.cons_index & ring->size_mask;
744 	cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
745 
746 	/* Process all completed CQEs */
747 	while (XNOR(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK,
748 		    cq->mcq.cons_index & cq->size)) {
749 		struct mlx4_en_rx_alloc *frags;
750 		enum pkt_hash_types hash_type;
751 		struct sk_buff *skb;
752 		unsigned int length;
753 		int ip_summed;
754 		void *va;
755 		int nr;
756 
757 		frags = ring->rx_info + (index << priv->log_rx_info);
758 		va = page_address(frags[0].page) + frags[0].page_offset;
759 		net_prefetchw(va);
760 		/*
761 		 * make sure we read the CQE after we read the ownership bit
762 		 */
763 		dma_rmb();
764 
765 		/* Drop packet on bad receive or bad checksum */
766 		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
767 						MLX4_CQE_OPCODE_ERROR)) {
768 			en_err(priv, "CQE completed in error - vendor syndrome:%d syndrome:%d\n",
769 			       ((struct mlx4_err_cqe *)cqe)->vendor_err_syndrome,
770 			       ((struct mlx4_err_cqe *)cqe)->syndrome);
771 			goto next;
772 		}
773 		if (unlikely(cqe->badfcs_enc & MLX4_CQE_BAD_FCS)) {
774 			en_dbg(RX_ERR, priv, "Accepted frame with bad FCS\n");
775 			goto next;
776 		}
777 
778 		/* Check if we need to drop the packet if SRIOV is not enabled
779 		 * and not performing the selftest or flb disabled
780 		 */
781 		if (priv->flags & MLX4_EN_FLAG_RX_FILTER_NEEDED) {
782 			const struct ethhdr *ethh = va;
783 			dma_addr_t dma;
784 			/* Get pointer to first fragment since we haven't
785 			 * skb yet and cast it to ethhdr struct
786 			 */
787 			dma = frags[0].dma + frags[0].page_offset;
788 			dma_sync_single_for_cpu(priv->ddev, dma, sizeof(*ethh),
789 						DMA_FROM_DEVICE);
790 
791 			if (is_multicast_ether_addr(ethh->h_dest)) {
792 				struct mlx4_mac_entry *entry;
793 				struct hlist_head *bucket;
794 				unsigned int mac_hash;
795 
796 				/* Drop the packet, since HW loopback-ed it */
797 				mac_hash = ethh->h_source[MLX4_EN_MAC_HASH_IDX];
798 				bucket = &priv->mac_hash[mac_hash];
799 				hlist_for_each_entry_rcu_bh(entry, bucket, hlist) {
800 					if (ether_addr_equal_64bits(entry->mac,
801 								    ethh->h_source))
802 						goto next;
803 				}
804 			}
805 		}
806 
807 		if (unlikely(priv->validate_loopback)) {
808 			validate_loopback(priv, va);
809 			goto next;
810 		}
811 
812 		/*
813 		 * Packet is OK - process it.
814 		 */
815 		length = be32_to_cpu(cqe->byte_cnt);
816 		length -= ring->fcs_del;
817 
818 		/* A bpf program gets first chance to drop the packet. It may
819 		 * read bytes but not past the end of the frag.
820 		 */
821 		if (xdp_prog) {
822 			dma_addr_t dma;
823 			void *orig_data;
824 			u32 act;
825 
826 			dma = frags[0].dma + frags[0].page_offset;
827 			dma_sync_single_for_cpu(priv->ddev, dma,
828 						priv->frag_info[0].frag_size,
829 						DMA_FROM_DEVICE);
830 
831 			xdp_prepare_buff(&mxbuf.xdp, va - frags[0].page_offset,
832 					 frags[0].page_offset, length, true);
833 			orig_data = mxbuf.xdp.data;
834 			mxbuf.cqe = cqe;
835 			mxbuf.mdev = priv->mdev;
836 			mxbuf.ring = ring;
837 			mxbuf.dev = dev;
838 
839 			act = bpf_prog_run_xdp(xdp_prog, &mxbuf.xdp);
840 
841 			length = mxbuf.xdp.data_end - mxbuf.xdp.data;
842 			if (mxbuf.xdp.data != orig_data) {
843 				frags[0].page_offset = mxbuf.xdp.data -
844 					mxbuf.xdp.data_hard_start;
845 				va = mxbuf.xdp.data;
846 			}
847 
848 			switch (act) {
849 			case XDP_PASS:
850 				break;
851 			case XDP_REDIRECT:
852 				if (likely(!xdp_do_redirect(dev, &mxbuf.xdp, xdp_prog))) {
853 					ring->xdp_redirect++;
854 					xdp_redir_flush = true;
855 					frags[0].page = NULL;
856 					goto next;
857 				}
858 				ring->xdp_redirect_fail++;
859 				trace_xdp_exception(dev, xdp_prog, act);
860 				goto xdp_drop_no_cnt;
861 			case XDP_TX:
862 				if (likely(!mlx4_en_xmit_frame(ring, frags, priv,
863 							length, cq_ring,
864 							&doorbell_pending))) {
865 					frags[0].page = NULL;
866 					goto next;
867 				}
868 				trace_xdp_exception(dev, xdp_prog, act);
869 				goto xdp_drop_no_cnt; /* Drop on xmit failure */
870 			default:
871 				bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
872 				fallthrough;
873 			case XDP_ABORTED:
874 				trace_xdp_exception(dev, xdp_prog, act);
875 				fallthrough;
876 			case XDP_DROP:
877 				ring->xdp_drop++;
878 xdp_drop_no_cnt:
879 				goto next;
880 			}
881 		}
882 
883 		ring->bytes += length;
884 		ring->packets++;
885 
886 		skb = napi_get_frags(&cq->napi);
887 		if (unlikely(!skb))
888 			goto next;
889 
890 		if (unlikely(ring->hwtstamp_rx_filter == HWTSTAMP_FILTER_ALL)) {
891 			u64 timestamp = mlx4_en_get_cqe_ts(cqe);
892 
893 			mlx4_en_fill_hwtstamps(priv->mdev, skb_hwtstamps(skb),
894 					       timestamp);
895 		}
896 		skb_record_rx_queue(skb, cq_ring);
897 
898 		if (likely(dev->features & NETIF_F_RXCSUM)) {
899 			/* TODO: For IP non TCP/UDP packets when csum complete is
900 			 * not an option (not supported or any other reason) we can
901 			 * actually check cqe IPOK status bit and report
902 			 * CHECKSUM_UNNECESSARY rather than CHECKSUM_NONE
903 			 */
904 			if ((cqe->status & cpu_to_be16(MLX4_CQE_STATUS_TCP |
905 						       MLX4_CQE_STATUS_UDP)) &&
906 			    (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IPOK)) &&
907 			    cqe->checksum == cpu_to_be16(0xffff)) {
908 				bool l2_tunnel;
909 
910 				l2_tunnel = (dev->hw_enc_features & NETIF_F_RXCSUM) &&
911 					(cqe->vlan_my_qpn & cpu_to_be32(MLX4_CQE_L2_TUNNEL));
912 				ip_summed = CHECKSUM_UNNECESSARY;
913 				hash_type = PKT_HASH_TYPE_L4;
914 				if (l2_tunnel)
915 					skb->csum_level = 1;
916 				ring->csum_ok++;
917 			} else {
918 				if (!(priv->flags & MLX4_EN_FLAG_RX_CSUM_NON_TCP_UDP &&
919 				      (cqe->status & cpu_to_be16(MLX4_CQE_STATUS_IP_ANY))))
920 					goto csum_none;
921 				if (check_csum(cqe, skb, va, dev->features))
922 					goto csum_none;
923 				ip_summed = CHECKSUM_COMPLETE;
924 				hash_type = PKT_HASH_TYPE_L3;
925 				ring->csum_complete++;
926 			}
927 		} else {
928 csum_none:
929 			ip_summed = CHECKSUM_NONE;
930 			hash_type = PKT_HASH_TYPE_L3;
931 			ring->csum_none++;
932 		}
933 		skb->ip_summed = ip_summed;
934 		if (dev->features & NETIF_F_RXHASH)
935 			skb_set_hash(skb,
936 				     be32_to_cpu(cqe->immed_rss_invalid),
937 				     hash_type);
938 
939 		if ((cqe->vlan_my_qpn &
940 		     cpu_to_be32(MLX4_CQE_CVLAN_PRESENT_MASK)) &&
941 		    (dev->features & NETIF_F_HW_VLAN_CTAG_RX))
942 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
943 					       be16_to_cpu(cqe->sl_vid));
944 		else if ((cqe->vlan_my_qpn &
945 			  cpu_to_be32(MLX4_CQE_SVLAN_PRESENT_MASK)) &&
946 			 (dev->features & NETIF_F_HW_VLAN_STAG_RX))
947 			__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021AD),
948 					       be16_to_cpu(cqe->sl_vid));
949 
950 		nr = mlx4_en_complete_rx_desc(priv, frags, skb, length);
951 		if (likely(nr)) {
952 			skb_shinfo(skb)->nr_frags = nr;
953 			skb->len = length;
954 			skb->data_len = length;
955 			napi_gro_frags(&cq->napi);
956 		} else {
957 			__vlan_hwaccel_clear_tag(skb);
958 			skb_clear_hash(skb);
959 		}
960 next:
961 		++cq->mcq.cons_index;
962 		index = (cq->mcq.cons_index) & ring->size_mask;
963 		cqe = mlx4_en_get_cqe(cq->buf, index, priv->cqe_size) + factor;
964 		if (unlikely(++polled == budget))
965 			break;
966 	}
967 
968 	if (xdp_redir_flush)
969 		xdp_do_flush();
970 
971 	if (likely(polled)) {
972 		if (doorbell_pending) {
973 			priv->tx_cq[TX_XDP][cq_ring]->xdp_busy = true;
974 			mlx4_en_xmit_doorbell(priv->tx_ring[TX_XDP][cq_ring]);
975 		}
976 
977 		mlx4_cq_set_ci(&cq->mcq);
978 		wmb(); /* ensure HW sees CQ consumer before we post new buffers */
979 		ring->cons = cq->mcq.cons_index;
980 	}
981 
982 	mlx4_en_refill_rx_buffers(priv, ring);
983 
984 	return polled;
985 }
986 
987 
988 void mlx4_en_rx_irq(struct mlx4_cq *mcq)
989 {
990 	struct mlx4_en_cq *cq = container_of(mcq, struct mlx4_en_cq, mcq);
991 	struct mlx4_en_priv *priv = netdev_priv(cq->dev);
992 
993 	if (likely(priv->port_up))
994 		napi_schedule_irqoff(&cq->napi);
995 	else
996 		mlx4_en_arm_cq(priv, cq);
997 }
998 
999 /* Rx CQ polling - called by NAPI */
1000 int mlx4_en_poll_rx_cq(struct napi_struct *napi, int budget)
1001 {
1002 	struct mlx4_en_cq *cq = container_of(napi, struct mlx4_en_cq, napi);
1003 	struct net_device *dev = cq->dev;
1004 	struct mlx4_en_priv *priv = netdev_priv(dev);
1005 	struct mlx4_en_cq *xdp_tx_cq = NULL;
1006 	bool clean_complete = true;
1007 	int done;
1008 
1009 	if (!budget)
1010 		return 0;
1011 
1012 	if (priv->tx_ring_num[TX_XDP]) {
1013 		xdp_tx_cq = priv->tx_cq[TX_XDP][cq->ring];
1014 		if (xdp_tx_cq->xdp_busy) {
1015 			clean_complete = mlx4_en_process_tx_cq(dev, xdp_tx_cq,
1016 							       budget) < budget;
1017 			xdp_tx_cq->xdp_busy = !clean_complete;
1018 		}
1019 	}
1020 
1021 	done = mlx4_en_process_rx_cq(dev, cq, budget);
1022 
1023 	/* If we used up all the quota - we're probably not done yet... */
1024 	if (done == budget || !clean_complete) {
1025 		int cpu_curr;
1026 
1027 		/* in case we got here because of !clean_complete */
1028 		done = budget;
1029 
1030 		cpu_curr = smp_processor_id();
1031 
1032 		if (likely(cpumask_test_cpu(cpu_curr, cq->aff_mask)))
1033 			return budget;
1034 
1035 		/* Current cpu is not according to smp_irq_affinity -
1036 		 * probably affinity changed. Need to stop this NAPI
1037 		 * poll, and restart it on the right CPU.
1038 		 * Try to avoid returning a too small value (like 0),
1039 		 * to not fool net_rx_action() and its netdev_budget
1040 		 */
1041 		if (done)
1042 			done--;
1043 	}
1044 	/* Done for now */
1045 	if (likely(napi_complete_done(napi, done)))
1046 		mlx4_en_arm_cq(priv, cq);
1047 	return done;
1048 }
1049 
1050 void mlx4_en_calc_rx_buf(struct net_device *dev)
1051 {
1052 	struct mlx4_en_priv *priv = netdev_priv(dev);
1053 	int eff_mtu = MLX4_EN_EFF_MTU(dev->mtu);
1054 	int i = 0;
1055 
1056 	/* bpf requires buffers to be set up as 1 packet per page.
1057 	 * This only works when num_frags == 1.
1058 	 */
1059 	if (priv->tx_ring_num[TX_XDP]) {
1060 		priv->frag_info[0].frag_size = eff_mtu;
1061 		/* This will gain efficient xdp frame recycling at the
1062 		 * expense of more costly truesize accounting
1063 		 */
1064 		priv->frag_info[0].frag_stride = PAGE_SIZE;
1065 		priv->dma_dir = DMA_BIDIRECTIONAL;
1066 		priv->rx_headroom = XDP_PACKET_HEADROOM;
1067 		i = 1;
1068 	} else {
1069 		int frag_size_max = 2048, buf_size = 0;
1070 
1071 		/* should not happen, right ? */
1072 		if (eff_mtu > PAGE_SIZE + (MLX4_EN_MAX_RX_FRAGS - 1) * 2048)
1073 			frag_size_max = PAGE_SIZE;
1074 
1075 		while (buf_size < eff_mtu) {
1076 			int frag_stride, frag_size = eff_mtu - buf_size;
1077 			int pad, nb;
1078 
1079 			if (i < MLX4_EN_MAX_RX_FRAGS - 1)
1080 				frag_size = min(frag_size, frag_size_max);
1081 
1082 			priv->frag_info[i].frag_size = frag_size;
1083 			frag_stride = ALIGN(frag_size, SMP_CACHE_BYTES);
1084 			/* We can only pack 2 1536-bytes frames in on 4K page
1085 			 * Therefore, each frame would consume more bytes (truesize)
1086 			 */
1087 			nb = PAGE_SIZE / frag_stride;
1088 			pad = (PAGE_SIZE - nb * frag_stride) / nb;
1089 			pad &= ~(SMP_CACHE_BYTES - 1);
1090 			priv->frag_info[i].frag_stride = frag_stride + pad;
1091 
1092 			buf_size += frag_size;
1093 			i++;
1094 		}
1095 		priv->dma_dir = DMA_FROM_DEVICE;
1096 		priv->rx_headroom = 0;
1097 	}
1098 
1099 	priv->num_frags = i;
1100 	priv->rx_skb_size = eff_mtu;
1101 	priv->log_rx_info = ROUNDUP_LOG2(i * sizeof(struct mlx4_en_rx_alloc));
1102 
1103 	en_dbg(DRV, priv, "Rx buffer scatter-list (effective-mtu:%d num_frags:%d):\n",
1104 	       eff_mtu, priv->num_frags);
1105 	for (i = 0; i < priv->num_frags; i++) {
1106 		en_dbg(DRV,
1107 		       priv,
1108 		       "  frag:%d - size:%d stride:%d\n",
1109 		       i,
1110 		       priv->frag_info[i].frag_size,
1111 		       priv->frag_info[i].frag_stride);
1112 	}
1113 }
1114 
1115 /* RSS related functions */
1116 
1117 static int mlx4_en_config_rss_qp(struct mlx4_en_priv *priv, int qpn,
1118 				 struct mlx4_en_rx_ring *ring,
1119 				 enum mlx4_qp_state *state,
1120 				 struct mlx4_qp *qp)
1121 {
1122 	struct mlx4_en_dev *mdev = priv->mdev;
1123 	struct mlx4_qp_context *context;
1124 	int err = 0;
1125 
1126 	context = kzalloc(sizeof(*context), GFP_KERNEL);
1127 	if (!context)
1128 		return -ENOMEM;
1129 
1130 	err = mlx4_qp_alloc(mdev->dev, qpn, qp);
1131 	if (err) {
1132 		en_err(priv, "Failed to allocate qp #%x\n", qpn);
1133 		goto out;
1134 	}
1135 	qp->event = mlx4_en_sqp_event;
1136 
1137 	mlx4_en_fill_qp_context(priv, ring->actual_size, ring->stride, 0, 0,
1138 				qpn, ring->cqn, -1, context);
1139 	context->db_rec_addr = cpu_to_be64(ring->wqres.db.dma);
1140 
1141 	/* Cancel FCS removal if FW allows */
1142 	if (mdev->dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP) {
1143 		context->param3 |= cpu_to_be32(1 << 29);
1144 		if (priv->dev->features & NETIF_F_RXFCS)
1145 			ring->fcs_del = 0;
1146 		else
1147 			ring->fcs_del = ETH_FCS_LEN;
1148 	} else
1149 		ring->fcs_del = 0;
1150 
1151 	err = mlx4_qp_to_ready(mdev->dev, &ring->wqres.mtt, context, qp, state);
1152 	if (err) {
1153 		mlx4_qp_remove(mdev->dev, qp);
1154 		mlx4_qp_free(mdev->dev, qp);
1155 	}
1156 	mlx4_en_update_rx_prod_db(ring);
1157 out:
1158 	kfree(context);
1159 	return err;
1160 }
1161 
1162 int mlx4_en_create_drop_qp(struct mlx4_en_priv *priv)
1163 {
1164 	int err;
1165 	u32 qpn;
1166 
1167 	err = mlx4_qp_reserve_range(priv->mdev->dev, 1, 1, &qpn,
1168 				    MLX4_RESERVE_A0_QP,
1169 				    MLX4_RES_USAGE_DRIVER);
1170 	if (err) {
1171 		en_err(priv, "Failed reserving drop qpn\n");
1172 		return err;
1173 	}
1174 	err = mlx4_qp_alloc(priv->mdev->dev, qpn, &priv->drop_qp);
1175 	if (err) {
1176 		en_err(priv, "Failed allocating drop qp\n");
1177 		mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1178 		return err;
1179 	}
1180 
1181 	return 0;
1182 }
1183 
1184 void mlx4_en_destroy_drop_qp(struct mlx4_en_priv *priv)
1185 {
1186 	u32 qpn;
1187 
1188 	qpn = priv->drop_qp.qpn;
1189 	mlx4_qp_remove(priv->mdev->dev, &priv->drop_qp);
1190 	mlx4_qp_free(priv->mdev->dev, &priv->drop_qp);
1191 	mlx4_qp_release_range(priv->mdev->dev, qpn, 1);
1192 }
1193 
1194 /* Allocate rx qp's and configure them according to rss map */
1195 int mlx4_en_config_rss_steer(struct mlx4_en_priv *priv)
1196 {
1197 	struct mlx4_en_dev *mdev = priv->mdev;
1198 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1199 	struct mlx4_qp_context context;
1200 	struct mlx4_rss_context *rss_context;
1201 	int rss_rings;
1202 	void *ptr;
1203 	u8 rss_mask = (MLX4_RSS_IPV4 | MLX4_RSS_TCP_IPV4 | MLX4_RSS_IPV6 |
1204 			MLX4_RSS_TCP_IPV6);
1205 	int i, qpn;
1206 	int err = 0;
1207 	int good_qps = 0;
1208 	u8 flags;
1209 
1210 	en_dbg(DRV, priv, "Configuring rss steering\n");
1211 
1212 	flags = priv->rx_ring_num == 1 ? MLX4_RESERVE_A0_QP : 0;
1213 	err = mlx4_qp_reserve_range(mdev->dev, priv->rx_ring_num,
1214 				    priv->rx_ring_num,
1215 				    &rss_map->base_qpn, flags,
1216 				    MLX4_RES_USAGE_DRIVER);
1217 	if (err) {
1218 		en_err(priv, "Failed reserving %d qps\n", priv->rx_ring_num);
1219 		return err;
1220 	}
1221 
1222 	for (i = 0; i < priv->rx_ring_num; i++) {
1223 		qpn = rss_map->base_qpn + i;
1224 		err = mlx4_en_config_rss_qp(priv, qpn, priv->rx_ring[i],
1225 					    &rss_map->state[i],
1226 					    &rss_map->qps[i]);
1227 		if (err)
1228 			goto rss_err;
1229 
1230 		++good_qps;
1231 	}
1232 
1233 	if (priv->rx_ring_num == 1) {
1234 		rss_map->indir_qp = &rss_map->qps[0];
1235 		priv->base_qpn = rss_map->indir_qp->qpn;
1236 		en_info(priv, "Optimized Non-RSS steering\n");
1237 		return 0;
1238 	}
1239 
1240 	rss_map->indir_qp = kzalloc(sizeof(*rss_map->indir_qp), GFP_KERNEL);
1241 	if (!rss_map->indir_qp) {
1242 		err = -ENOMEM;
1243 		goto rss_err;
1244 	}
1245 
1246 	/* Configure RSS indirection qp */
1247 	err = mlx4_qp_alloc(mdev->dev, priv->base_qpn, rss_map->indir_qp);
1248 	if (err) {
1249 		en_err(priv, "Failed to allocate RSS indirection QP\n");
1250 		goto qp_alloc_err;
1251 	}
1252 
1253 	rss_map->indir_qp->event = mlx4_en_sqp_event;
1254 	mlx4_en_fill_qp_context(priv, 0, 0, 0, 1, priv->base_qpn,
1255 				priv->rx_ring[0]->cqn, -1, &context);
1256 
1257 	if (!priv->prof->rss_rings || priv->prof->rss_rings > priv->rx_ring_num)
1258 		rss_rings = priv->rx_ring_num;
1259 	else
1260 		rss_rings = priv->prof->rss_rings;
1261 
1262 	ptr = ((void *) &context) + offsetof(struct mlx4_qp_context, pri_path)
1263 					+ MLX4_RSS_OFFSET_IN_QPC_PRI_PATH;
1264 	rss_context = ptr;
1265 	rss_context->base_qpn = cpu_to_be32(ilog2(rss_rings) << 24 |
1266 					    (rss_map->base_qpn));
1267 	rss_context->default_qpn = cpu_to_be32(rss_map->base_qpn);
1268 	if (priv->mdev->profile.udp_rss) {
1269 		rss_mask |=  MLX4_RSS_UDP_IPV4 | MLX4_RSS_UDP_IPV6;
1270 		rss_context->base_qpn_udp = rss_context->default_qpn;
1271 	}
1272 
1273 	if (mdev->dev->caps.tunnel_offload_mode == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) {
1274 		en_info(priv, "Setting RSS context tunnel type to RSS on inner headers\n");
1275 		rss_mask |= MLX4_RSS_BY_INNER_HEADERS;
1276 	}
1277 
1278 	rss_context->flags = rss_mask;
1279 	rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1280 	if (priv->rss_hash_fn == ETH_RSS_HASH_XOR) {
1281 		rss_context->hash_fn = MLX4_RSS_HASH_XOR;
1282 	} else if (priv->rss_hash_fn == ETH_RSS_HASH_TOP) {
1283 		rss_context->hash_fn = MLX4_RSS_HASH_TOP;
1284 		memcpy(rss_context->rss_key, priv->rss_key,
1285 		       MLX4_EN_RSS_KEY_SIZE);
1286 	} else {
1287 		en_err(priv, "Unknown RSS hash function requested\n");
1288 		err = -EINVAL;
1289 		goto indir_err;
1290 	}
1291 
1292 	err = mlx4_qp_to_ready(mdev->dev, &priv->res.mtt, &context,
1293 			       rss_map->indir_qp, &rss_map->indir_state);
1294 	if (err)
1295 		goto indir_err;
1296 
1297 	return 0;
1298 
1299 indir_err:
1300 	mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1301 		       MLX4_QP_STATE_RST, NULL, 0, 0, rss_map->indir_qp);
1302 	mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
1303 	mlx4_qp_free(mdev->dev, rss_map->indir_qp);
1304 qp_alloc_err:
1305 	kfree(rss_map->indir_qp);
1306 	rss_map->indir_qp = NULL;
1307 rss_err:
1308 	for (i = 0; i < good_qps; i++) {
1309 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1310 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1311 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1312 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1313 	}
1314 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1315 	return err;
1316 }
1317 
1318 void mlx4_en_release_rss_steer(struct mlx4_en_priv *priv)
1319 {
1320 	struct mlx4_en_dev *mdev = priv->mdev;
1321 	struct mlx4_en_rss_map *rss_map = &priv->rss_map;
1322 	int i;
1323 
1324 	if (priv->rx_ring_num > 1) {
1325 		mlx4_qp_modify(mdev->dev, NULL, rss_map->indir_state,
1326 			       MLX4_QP_STATE_RST, NULL, 0, 0,
1327 			       rss_map->indir_qp);
1328 		mlx4_qp_remove(mdev->dev, rss_map->indir_qp);
1329 		mlx4_qp_free(mdev->dev, rss_map->indir_qp);
1330 		kfree(rss_map->indir_qp);
1331 		rss_map->indir_qp = NULL;
1332 	}
1333 
1334 	for (i = 0; i < priv->rx_ring_num; i++) {
1335 		mlx4_qp_modify(mdev->dev, NULL, rss_map->state[i],
1336 			       MLX4_QP_STATE_RST, NULL, 0, 0, &rss_map->qps[i]);
1337 		mlx4_qp_remove(mdev->dev, &rss_map->qps[i]);
1338 		mlx4_qp_free(mdev->dev, &rss_map->qps[i]);
1339 	}
1340 	mlx4_qp_release_range(mdev->dev, rss_map->base_qpn, priv->rx_ring_num);
1341 }
1342