1 /*
2  * Copyright (c) 2015, 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 #include <linux/prefetch.h>
34 #include <linux/ip.h>
35 #include <linux/ipv6.h>
36 #include <linux/tcp.h>
37 #include <linux/bpf_trace.h>
38 #include <net/busy_poll.h>
39 #include "en.h"
40 #include "en_tc.h"
41 #include "eswitch.h"
42 #include "en_rep.h"
43 #include "ipoib/ipoib.h"
44 #include "en_accel/ipsec_rxtx.h"
45 
46 static inline bool mlx5e_rx_hw_stamp(struct mlx5e_tstamp *tstamp)
47 {
48 	return tstamp->hwtstamp_config.rx_filter == HWTSTAMP_FILTER_ALL;
49 }
50 
51 static inline void mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cqcc,
52 				       void *data)
53 {
54 	u32 ci = cqcc & cq->wq.sz_m1;
55 
56 	memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, ci), sizeof(struct mlx5_cqe64));
57 }
58 
59 static inline void mlx5e_read_title_slot(struct mlx5e_rq *rq,
60 					 struct mlx5e_cq *cq, u32 cqcc)
61 {
62 	mlx5e_read_cqe_slot(cq, cqcc, &cq->title);
63 	cq->decmprs_left        = be32_to_cpu(cq->title.byte_cnt);
64 	cq->decmprs_wqe_counter = be16_to_cpu(cq->title.wqe_counter);
65 	rq->stats.cqe_compress_blks++;
66 }
67 
68 static inline void mlx5e_read_mini_arr_slot(struct mlx5e_cq *cq, u32 cqcc)
69 {
70 	mlx5e_read_cqe_slot(cq, cqcc, cq->mini_arr);
71 	cq->mini_arr_idx = 0;
72 }
73 
74 static inline void mlx5e_cqes_update_owner(struct mlx5e_cq *cq, u32 cqcc, int n)
75 {
76 	u8 op_own = (cqcc >> cq->wq.log_sz) & 1;
77 	u32 wq_sz = 1 << cq->wq.log_sz;
78 	u32 ci = cqcc & cq->wq.sz_m1;
79 	u32 ci_top = min_t(u32, wq_sz, ci + n);
80 
81 	for (; ci < ci_top; ci++, n--) {
82 		struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
83 
84 		cqe->op_own = op_own;
85 	}
86 
87 	if (unlikely(ci == wq_sz)) {
88 		op_own = !op_own;
89 		for (ci = 0; ci < n; ci++) {
90 			struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
91 
92 			cqe->op_own = op_own;
93 		}
94 	}
95 }
96 
97 static inline void mlx5e_decompress_cqe(struct mlx5e_rq *rq,
98 					struct mlx5e_cq *cq, u32 cqcc)
99 {
100 	cq->title.byte_cnt     = cq->mini_arr[cq->mini_arr_idx].byte_cnt;
101 	cq->title.check_sum    = cq->mini_arr[cq->mini_arr_idx].checksum;
102 	cq->title.op_own      &= 0xf0;
103 	cq->title.op_own      |= 0x01 & (cqcc >> cq->wq.log_sz);
104 	cq->title.wqe_counter  = cpu_to_be16(cq->decmprs_wqe_counter);
105 
106 	if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
107 		cq->decmprs_wqe_counter +=
108 			mpwrq_get_cqe_consumed_strides(&cq->title);
109 	else
110 		cq->decmprs_wqe_counter =
111 			(cq->decmprs_wqe_counter + 1) & rq->wq.sz_m1;
112 }
113 
114 static inline void mlx5e_decompress_cqe_no_hash(struct mlx5e_rq *rq,
115 						struct mlx5e_cq *cq, u32 cqcc)
116 {
117 	mlx5e_decompress_cqe(rq, cq, cqcc);
118 	cq->title.rss_hash_type   = 0;
119 	cq->title.rss_hash_result = 0;
120 }
121 
122 static inline u32 mlx5e_decompress_cqes_cont(struct mlx5e_rq *rq,
123 					     struct mlx5e_cq *cq,
124 					     int update_owner_only,
125 					     int budget_rem)
126 {
127 	u32 cqcc = cq->wq.cc + update_owner_only;
128 	u32 cqe_count;
129 	u32 i;
130 
131 	cqe_count = min_t(u32, cq->decmprs_left, budget_rem);
132 
133 	for (i = update_owner_only; i < cqe_count;
134 	     i++, cq->mini_arr_idx++, cqcc++) {
135 		if (cq->mini_arr_idx == MLX5_MINI_CQE_ARRAY_SIZE)
136 			mlx5e_read_mini_arr_slot(cq, cqcc);
137 
138 		mlx5e_decompress_cqe_no_hash(rq, cq, cqcc);
139 		rq->handle_rx_cqe(rq, &cq->title);
140 	}
141 	mlx5e_cqes_update_owner(cq, cq->wq.cc, cqcc - cq->wq.cc);
142 	cq->wq.cc = cqcc;
143 	cq->decmprs_left -= cqe_count;
144 	rq->stats.cqe_compress_pkts += cqe_count;
145 
146 	return cqe_count;
147 }
148 
149 static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
150 					      struct mlx5e_cq *cq,
151 					      int budget_rem)
152 {
153 	mlx5e_read_title_slot(rq, cq, cq->wq.cc);
154 	mlx5e_read_mini_arr_slot(cq, cq->wq.cc + 1);
155 	mlx5e_decompress_cqe(rq, cq, cq->wq.cc);
156 	rq->handle_rx_cqe(rq, &cq->title);
157 	cq->mini_arr_idx++;
158 
159 	return mlx5e_decompress_cqes_cont(rq, cq, 1, budget_rem) - 1;
160 }
161 
162 #define RQ_PAGE_SIZE(rq) ((1 << rq->buff.page_order) << PAGE_SHIFT)
163 
164 static inline bool mlx5e_page_is_reserved(struct page *page)
165 {
166 	return page_is_pfmemalloc(page) || page_to_nid(page) != numa_mem_id();
167 }
168 
169 static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq,
170 				      struct mlx5e_dma_info *dma_info)
171 {
172 	struct mlx5e_page_cache *cache = &rq->page_cache;
173 	u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
174 
175 	if (tail_next == cache->head) {
176 		rq->stats.cache_full++;
177 		return false;
178 	}
179 
180 	if (unlikely(mlx5e_page_is_reserved(dma_info->page))) {
181 		rq->stats.cache_waive++;
182 		return false;
183 	}
184 
185 	cache->page_cache[cache->tail] = *dma_info;
186 	cache->tail = tail_next;
187 	return true;
188 }
189 
190 static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq,
191 				      struct mlx5e_dma_info *dma_info)
192 {
193 	struct mlx5e_page_cache *cache = &rq->page_cache;
194 
195 	if (unlikely(cache->head == cache->tail)) {
196 		rq->stats.cache_empty++;
197 		return false;
198 	}
199 
200 	if (page_ref_count(cache->page_cache[cache->head].page) != 1) {
201 		rq->stats.cache_busy++;
202 		return false;
203 	}
204 
205 	*dma_info = cache->page_cache[cache->head];
206 	cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
207 	rq->stats.cache_reuse++;
208 
209 	dma_sync_single_for_device(rq->pdev, dma_info->addr,
210 				   RQ_PAGE_SIZE(rq),
211 				   DMA_FROM_DEVICE);
212 	return true;
213 }
214 
215 static inline int mlx5e_page_alloc_mapped(struct mlx5e_rq *rq,
216 					  struct mlx5e_dma_info *dma_info)
217 {
218 	struct page *page;
219 
220 	if (mlx5e_rx_cache_get(rq, dma_info))
221 		return 0;
222 
223 	page = dev_alloc_pages(rq->buff.page_order);
224 	if (unlikely(!page))
225 		return -ENOMEM;
226 
227 	dma_info->addr = dma_map_page(rq->pdev, page, 0,
228 				      RQ_PAGE_SIZE(rq), rq->buff.map_dir);
229 	if (unlikely(dma_mapping_error(rq->pdev, dma_info->addr))) {
230 		put_page(page);
231 		return -ENOMEM;
232 	}
233 	dma_info->page = page;
234 
235 	return 0;
236 }
237 
238 void mlx5e_page_release(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info,
239 			bool recycle)
240 {
241 	if (likely(recycle) && mlx5e_rx_cache_put(rq, dma_info))
242 		return;
243 
244 	dma_unmap_page(rq->pdev, dma_info->addr, RQ_PAGE_SIZE(rq),
245 		       rq->buff.map_dir);
246 	put_page(dma_info->page);
247 }
248 
249 static inline bool mlx5e_page_reuse(struct mlx5e_rq *rq,
250 				    struct mlx5e_wqe_frag_info *wi)
251 {
252 	return rq->wqe.page_reuse && wi->di.page &&
253 		(wi->offset + rq->wqe.frag_sz <= RQ_PAGE_SIZE(rq)) &&
254 		!mlx5e_page_is_reserved(wi->di.page);
255 }
256 
257 static int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe *wqe, u16 ix)
258 {
259 	struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
260 
261 	/* check if page exists, hence can be reused */
262 	if (!wi->di.page) {
263 		if (unlikely(mlx5e_page_alloc_mapped(rq, &wi->di)))
264 			return -ENOMEM;
265 		wi->offset = 0;
266 	}
267 
268 	wqe->data.addr = cpu_to_be64(wi->di.addr + wi->offset + rq->buff.headroom);
269 	return 0;
270 }
271 
272 static inline void mlx5e_free_rx_wqe(struct mlx5e_rq *rq,
273 				     struct mlx5e_wqe_frag_info *wi)
274 {
275 	mlx5e_page_release(rq, &wi->di, true);
276 	wi->di.page = NULL;
277 }
278 
279 static inline void mlx5e_free_rx_wqe_reuse(struct mlx5e_rq *rq,
280 					   struct mlx5e_wqe_frag_info *wi)
281 {
282 	if (mlx5e_page_reuse(rq, wi)) {
283 		rq->stats.page_reuse++;
284 		return;
285 	}
286 
287 	mlx5e_free_rx_wqe(rq, wi);
288 }
289 
290 void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
291 {
292 	struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
293 
294 	if (wi->di.page)
295 		mlx5e_free_rx_wqe(rq, wi);
296 }
297 
298 static inline int mlx5e_mpwqe_strides_per_page(struct mlx5e_rq *rq)
299 {
300 	return rq->mpwqe.num_strides >> MLX5_MPWRQ_WQE_PAGE_ORDER;
301 }
302 
303 static inline void mlx5e_add_skb_frag_mpwqe(struct mlx5e_rq *rq,
304 					    struct sk_buff *skb,
305 					    struct mlx5e_mpw_info *wi,
306 					    u32 page_idx, u32 frag_offset,
307 					    u32 len)
308 {
309 	unsigned int truesize = ALIGN(len, BIT(rq->mpwqe.log_stride_sz));
310 
311 	dma_sync_single_for_cpu(rq->pdev,
312 				wi->umr.dma_info[page_idx].addr + frag_offset,
313 				len, DMA_FROM_DEVICE);
314 	wi->skbs_frags[page_idx]++;
315 	skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
316 			wi->umr.dma_info[page_idx].page, frag_offset,
317 			len, truesize);
318 }
319 
320 static inline void
321 mlx5e_copy_skb_header_mpwqe(struct device *pdev,
322 			    struct sk_buff *skb,
323 			    struct mlx5e_mpw_info *wi,
324 			    u32 page_idx, u32 offset,
325 			    u32 headlen)
326 {
327 	u16 headlen_pg = min_t(u32, headlen, PAGE_SIZE - offset);
328 	struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[page_idx];
329 	unsigned int len;
330 
331 	 /* Aligning len to sizeof(long) optimizes memcpy performance */
332 	len = ALIGN(headlen_pg, sizeof(long));
333 	dma_sync_single_for_cpu(pdev, dma_info->addr + offset, len,
334 				DMA_FROM_DEVICE);
335 	skb_copy_to_linear_data_offset(skb, 0,
336 				       page_address(dma_info->page) + offset,
337 				       len);
338 	if (unlikely(offset + headlen > PAGE_SIZE)) {
339 		dma_info++;
340 		headlen_pg = len;
341 		len = ALIGN(headlen - headlen_pg, sizeof(long));
342 		dma_sync_single_for_cpu(pdev, dma_info->addr, len,
343 					DMA_FROM_DEVICE);
344 		skb_copy_to_linear_data_offset(skb, headlen_pg,
345 					       page_address(dma_info->page),
346 					       len);
347 	}
348 }
349 
350 static inline void mlx5e_post_umr_wqe(struct mlx5e_rq *rq, u16 ix)
351 {
352 	struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
353 	struct mlx5e_icosq *sq = &rq->channel->icosq;
354 	struct mlx5_wq_cyc *wq = &sq->wq;
355 	struct mlx5e_umr_wqe *wqe;
356 	u8 num_wqebbs = DIV_ROUND_UP(sizeof(*wqe), MLX5_SEND_WQE_BB);
357 	u16 pi;
358 
359 	/* fill sq edge with nops to avoid wqe wrap around */
360 	while ((pi = (sq->pc & wq->sz_m1)) > sq->edge) {
361 		sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_NOP;
362 		mlx5e_post_nop(wq, sq->sqn, &sq->pc);
363 	}
364 
365 	wqe = mlx5_wq_cyc_get_wqe(wq, pi);
366 	memcpy(wqe, &wi->umr.wqe, sizeof(*wqe));
367 	wqe->ctrl.opmod_idx_opcode =
368 		cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
369 			    MLX5_OPCODE_UMR);
370 
371 	sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_UMR;
372 	sq->pc += num_wqebbs;
373 	mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &wqe->ctrl);
374 }
375 
376 static int mlx5e_alloc_rx_umr_mpwqe(struct mlx5e_rq *rq,
377 				    u16 ix)
378 {
379 	struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
380 	int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
381 	struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
382 	int err;
383 	int i;
384 
385 	for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++, dma_info++) {
386 		err = mlx5e_page_alloc_mapped(rq, dma_info);
387 		if (unlikely(err))
388 			goto err_unmap;
389 		wi->umr.mtt[i] = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
390 		page_ref_add(dma_info->page, pg_strides);
391 	}
392 
393 	memset(wi->skbs_frags, 0, sizeof(*wi->skbs_frags) * MLX5_MPWRQ_PAGES_PER_WQE);
394 	wi->consumed_strides = 0;
395 
396 	return 0;
397 
398 err_unmap:
399 	while (--i >= 0) {
400 		dma_info--;
401 		page_ref_sub(dma_info->page, pg_strides);
402 		mlx5e_page_release(rq, dma_info, true);
403 	}
404 
405 	return err;
406 }
407 
408 void mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi)
409 {
410 	int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
411 	struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
412 	int i;
413 
414 	for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++, dma_info++) {
415 		page_ref_sub(dma_info->page, pg_strides - wi->skbs_frags[i]);
416 		mlx5e_page_release(rq, dma_info, true);
417 	}
418 }
419 
420 static void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
421 {
422 	struct mlx5_wq_ll *wq = &rq->wq;
423 	struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
424 
425 	rq->mpwqe.umr_in_progress = false;
426 
427 	mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
428 
429 	/* ensure wqes are visible to device before updating doorbell record */
430 	dma_wmb();
431 
432 	mlx5_wq_ll_update_db_record(wq);
433 }
434 
435 static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
436 {
437 	int err;
438 
439 	err = mlx5e_alloc_rx_umr_mpwqe(rq, ix);
440 	if (unlikely(err)) {
441 		rq->stats.buff_alloc_err++;
442 		return err;
443 	}
444 	rq->mpwqe.umr_in_progress = true;
445 	mlx5e_post_umr_wqe(rq, ix);
446 	return 0;
447 }
448 
449 void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
450 {
451 	struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
452 
453 	mlx5e_free_rx_mpwqe(rq, wi);
454 }
455 
456 bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
457 {
458 	struct mlx5_wq_ll *wq = &rq->wq;
459 	int err;
460 
461 	if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
462 		return false;
463 
464 	if (mlx5_wq_ll_is_full(wq))
465 		return false;
466 
467 	do {
468 		struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
469 
470 		err = mlx5e_alloc_rx_wqe(rq, wqe, wq->head);
471 		if (unlikely(err)) {
472 			rq->stats.buff_alloc_err++;
473 			break;
474 		}
475 
476 		mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
477 	} while (!mlx5_wq_ll_is_full(wq));
478 
479 	/* ensure wqes are visible to device before updating doorbell record */
480 	dma_wmb();
481 
482 	mlx5_wq_ll_update_db_record(wq);
483 
484 	return !!err;
485 }
486 
487 static inline void mlx5e_poll_ico_single_cqe(struct mlx5e_cq *cq,
488 					     struct mlx5e_icosq *sq,
489 					     struct mlx5e_rq *rq,
490 					     struct mlx5_cqe64 *cqe)
491 {
492 	struct mlx5_wq_cyc *wq = &sq->wq;
493 	u16 ci = be16_to_cpu(cqe->wqe_counter) & wq->sz_m1;
494 	struct mlx5e_sq_wqe_info *icowi = &sq->db.ico_wqe[ci];
495 
496 	mlx5_cqwq_pop(&cq->wq);
497 
498 	if (unlikely((cqe->op_own >> 4) != MLX5_CQE_REQ)) {
499 		WARN_ONCE(true, "mlx5e: Bad OP in ICOSQ CQE: 0x%x\n",
500 			  cqe->op_own);
501 		return;
502 	}
503 
504 	if (likely(icowi->opcode == MLX5_OPCODE_UMR)) {
505 		mlx5e_post_rx_mpwqe(rq);
506 		return;
507 	}
508 
509 	if (unlikely(icowi->opcode != MLX5_OPCODE_NOP))
510 		WARN_ONCE(true,
511 			  "mlx5e: Bad OPCODE in ICOSQ WQE info: 0x%x\n",
512 			  icowi->opcode);
513 }
514 
515 static void mlx5e_poll_ico_cq(struct mlx5e_cq *cq, struct mlx5e_rq *rq)
516 {
517 	struct mlx5e_icosq *sq = container_of(cq, struct mlx5e_icosq, cq);
518 	struct mlx5_cqe64 *cqe;
519 
520 	if (unlikely(!MLX5E_TEST_BIT(sq->state, MLX5E_SQ_STATE_ENABLED)))
521 		return;
522 
523 	cqe = mlx5_cqwq_get_cqe(&cq->wq);
524 	if (likely(!cqe))
525 		return;
526 
527 	/* by design, there's only a single cqe */
528 	mlx5e_poll_ico_single_cqe(cq, sq, rq, cqe);
529 
530 	mlx5_cqwq_update_db_record(&cq->wq);
531 }
532 
533 bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq)
534 {
535 	struct mlx5_wq_ll *wq = &rq->wq;
536 
537 	if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
538 		return false;
539 
540 	mlx5e_poll_ico_cq(&rq->channel->icosq.cq, rq);
541 
542 	if (mlx5_wq_ll_is_full(wq))
543 		return false;
544 
545 	if (!rq->mpwqe.umr_in_progress)
546 		mlx5e_alloc_rx_mpwqe(rq, wq->head);
547 
548 	return true;
549 }
550 
551 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe,
552 				 u32 cqe_bcnt)
553 {
554 	struct ethhdr	*eth = (struct ethhdr *)(skb->data);
555 	struct tcphdr	*tcp;
556 	int network_depth = 0;
557 	__be16 proto;
558 	u16 tot_len;
559 	void *ip_p;
560 
561 	u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
562 	u8 tcp_ack = (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA) ||
563 		(l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA);
564 
565 	skb->mac_len = ETH_HLEN;
566 	proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
567 
568 	tot_len = cqe_bcnt - network_depth;
569 	ip_p = skb->data + network_depth;
570 
571 	if (proto == htons(ETH_P_IP)) {
572 		struct iphdr *ipv4 = ip_p;
573 
574 		tcp = ip_p + sizeof(struct iphdr);
575 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
576 
577 		ipv4->ttl               = cqe->lro_min_ttl;
578 		ipv4->tot_len           = cpu_to_be16(tot_len);
579 		ipv4->check             = 0;
580 		ipv4->check             = ip_fast_csum((unsigned char *)ipv4,
581 						       ipv4->ihl);
582 	} else {
583 		struct ipv6hdr *ipv6 = ip_p;
584 
585 		tcp = ip_p + sizeof(struct ipv6hdr);
586 		skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
587 
588 		ipv6->hop_limit         = cqe->lro_min_ttl;
589 		ipv6->payload_len       = cpu_to_be16(tot_len -
590 						      sizeof(struct ipv6hdr));
591 	}
592 
593 	tcp->psh = get_cqe_lro_tcppsh(cqe);
594 
595 	if (tcp_ack) {
596 		tcp->ack                = 1;
597 		tcp->ack_seq            = cqe->lro_ack_seq_num;
598 		tcp->window             = cqe->lro_tcp_win;
599 	}
600 }
601 
602 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
603 				      struct sk_buff *skb)
604 {
605 	u8 cht = cqe->rss_hash_type;
606 	int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
607 		 (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
608 					    PKT_HASH_TYPE_NONE;
609 	skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
610 }
611 
612 static inline bool is_first_ethertype_ip(struct sk_buff *skb)
613 {
614 	__be16 ethertype = ((struct ethhdr *)skb->data)->h_proto;
615 
616 	return (ethertype == htons(ETH_P_IP) || ethertype == htons(ETH_P_IPV6));
617 }
618 
619 static inline void mlx5e_handle_csum(struct net_device *netdev,
620 				     struct mlx5_cqe64 *cqe,
621 				     struct mlx5e_rq *rq,
622 				     struct sk_buff *skb,
623 				     bool   lro)
624 {
625 	if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
626 		goto csum_none;
627 
628 	if (lro) {
629 		skb->ip_summed = CHECKSUM_UNNECESSARY;
630 		rq->stats.csum_unnecessary++;
631 		return;
632 	}
633 
634 	if (is_first_ethertype_ip(skb)) {
635 		skb->ip_summed = CHECKSUM_COMPLETE;
636 		skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
637 		rq->stats.csum_complete++;
638 		return;
639 	}
640 
641 	if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
642 		   (cqe->hds_ip_ext & CQE_L4_OK))) {
643 		skb->ip_summed = CHECKSUM_UNNECESSARY;
644 		if (cqe_is_tunneled(cqe)) {
645 			skb->csum_level = 1;
646 			skb->encapsulation = 1;
647 			rq->stats.csum_unnecessary_inner++;
648 			return;
649 		}
650 		rq->stats.csum_unnecessary++;
651 		return;
652 	}
653 csum_none:
654 	skb->ip_summed = CHECKSUM_NONE;
655 	rq->stats.csum_none++;
656 }
657 
658 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
659 				      u32 cqe_bcnt,
660 				      struct mlx5e_rq *rq,
661 				      struct sk_buff *skb)
662 {
663 	struct net_device *netdev = rq->netdev;
664 	struct mlx5e_tstamp *tstamp = rq->tstamp;
665 	int lro_num_seg;
666 
667 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
668 	if (lro_num_seg > 1) {
669 		mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
670 		skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
671 		/* Subtract one since we already counted this as one
672 		 * "regular" packet in mlx5e_complete_rx_cqe()
673 		 */
674 		rq->stats.packets += lro_num_seg - 1;
675 		rq->stats.lro_packets++;
676 		rq->stats.lro_bytes += cqe_bcnt;
677 	}
678 
679 	if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
680 		mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
681 
682 	skb_record_rx_queue(skb, rq->ix);
683 
684 	if (likely(netdev->features & NETIF_F_RXHASH))
685 		mlx5e_skb_set_hash(cqe, skb);
686 
687 	if (cqe_has_vlan(cqe))
688 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
689 				       be16_to_cpu(cqe->vlan_info));
690 
691 	skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
692 
693 	mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
694 	skb->protocol = eth_type_trans(skb, netdev);
695 }
696 
697 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
698 					 struct mlx5_cqe64 *cqe,
699 					 u32 cqe_bcnt,
700 					 struct sk_buff *skb)
701 {
702 	rq->stats.packets++;
703 	rq->stats.bytes += cqe_bcnt;
704 	mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
705 }
706 
707 static inline void mlx5e_xmit_xdp_doorbell(struct mlx5e_xdpsq *sq)
708 {
709 	struct mlx5_wq_cyc *wq = &sq->wq;
710 	struct mlx5e_tx_wqe *wqe;
711 	u16 pi = (sq->pc - 1) & wq->sz_m1; /* last pi */
712 
713 	wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
714 
715 	mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &wqe->ctrl);
716 }
717 
718 static inline bool mlx5e_xmit_xdp_frame(struct mlx5e_rq *rq,
719 					struct mlx5e_dma_info *di,
720 					const struct xdp_buff *xdp)
721 {
722 	struct mlx5e_xdpsq       *sq   = &rq->xdpsq;
723 	struct mlx5_wq_cyc       *wq   = &sq->wq;
724 	u16                       pi   = sq->pc & wq->sz_m1;
725 	struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
726 
727 	struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
728 	struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
729 	struct mlx5_wqe_data_seg *dseg;
730 
731 	ptrdiff_t data_offset = xdp->data - xdp->data_hard_start;
732 	dma_addr_t dma_addr  = di->addr + data_offset;
733 	unsigned int dma_len = xdp->data_end - xdp->data;
734 
735 	prefetchw(wqe);
736 
737 	if (unlikely(dma_len < MLX5E_XDP_MIN_INLINE ||
738 		     MLX5E_SW2HW_MTU(rq->channel->priv, rq->netdev->mtu) < dma_len)) {
739 		rq->stats.xdp_drop++;
740 		return false;
741 	}
742 
743 	if (unlikely(!mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1))) {
744 		if (sq->db.doorbell) {
745 			/* SQ is full, ring doorbell */
746 			mlx5e_xmit_xdp_doorbell(sq);
747 			sq->db.doorbell = false;
748 		}
749 		rq->stats.xdp_tx_full++;
750 		return false;
751 	}
752 
753 	dma_sync_single_for_device(sq->pdev, dma_addr, dma_len, PCI_DMA_TODEVICE);
754 
755 	cseg->fm_ce_se = 0;
756 
757 	dseg = (struct mlx5_wqe_data_seg *)eseg + 1;
758 
759 	/* copy the inline part if required */
760 	if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
761 		memcpy(eseg->inline_hdr.start, xdp->data, MLX5E_XDP_MIN_INLINE);
762 		eseg->inline_hdr.sz = cpu_to_be16(MLX5E_XDP_MIN_INLINE);
763 		dma_len  -= MLX5E_XDP_MIN_INLINE;
764 		dma_addr += MLX5E_XDP_MIN_INLINE;
765 		dseg++;
766 	}
767 
768 	/* write the dma part */
769 	dseg->addr       = cpu_to_be64(dma_addr);
770 	dseg->byte_count = cpu_to_be32(dma_len);
771 
772 	cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_SEND);
773 
774 	/* move page to reference to sq responsibility,
775 	 * and mark so it's not put back in page-cache.
776 	 */
777 	rq->wqe.xdp_xmit = true;
778 	sq->db.di[pi] = *di;
779 	sq->pc++;
780 
781 	sq->db.doorbell = true;
782 
783 	rq->stats.xdp_tx++;
784 	return true;
785 }
786 
787 /* returns true if packet was consumed by xdp */
788 static inline int mlx5e_xdp_handle(struct mlx5e_rq *rq,
789 				   struct mlx5e_dma_info *di,
790 				   void *va, u16 *rx_headroom, u32 *len)
791 {
792 	const struct bpf_prog *prog = READ_ONCE(rq->xdp_prog);
793 	struct xdp_buff xdp;
794 	u32 act;
795 
796 	if (!prog)
797 		return false;
798 
799 	xdp.data = va + *rx_headroom;
800 	xdp.data_end = xdp.data + *len;
801 	xdp.data_hard_start = va;
802 
803 	act = bpf_prog_run_xdp(prog, &xdp);
804 	switch (act) {
805 	case XDP_PASS:
806 		*rx_headroom = xdp.data - xdp.data_hard_start;
807 		*len = xdp.data_end - xdp.data;
808 		return false;
809 	case XDP_TX:
810 		if (unlikely(!mlx5e_xmit_xdp_frame(rq, di, &xdp)))
811 			trace_xdp_exception(rq->netdev, prog, act);
812 		return true;
813 	default:
814 		bpf_warn_invalid_xdp_action(act);
815 	case XDP_ABORTED:
816 		trace_xdp_exception(rq->netdev, prog, act);
817 	case XDP_DROP:
818 		rq->stats.xdp_drop++;
819 		return true;
820 	}
821 }
822 
823 static inline
824 struct sk_buff *skb_from_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
825 			     struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
826 {
827 	struct mlx5e_dma_info *di = &wi->di;
828 	u16 rx_headroom = rq->buff.headroom;
829 	struct sk_buff *skb;
830 	void *va, *data;
831 	bool consumed;
832 	u32 frag_size;
833 
834 	va             = page_address(di->page) + wi->offset;
835 	data           = va + rx_headroom;
836 	frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
837 
838 	dma_sync_single_range_for_cpu(rq->pdev,
839 				      di->addr + wi->offset,
840 				      0, frag_size,
841 				      DMA_FROM_DEVICE);
842 	prefetch(data);
843 	wi->offset += frag_size;
844 
845 	if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
846 		rq->stats.wqe_err++;
847 		return NULL;
848 	}
849 
850 	rcu_read_lock();
851 	consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt);
852 	rcu_read_unlock();
853 	if (consumed)
854 		return NULL; /* page/packet was consumed by XDP */
855 
856 	skb = build_skb(va, frag_size);
857 	if (unlikely(!skb)) {
858 		rq->stats.buff_alloc_err++;
859 		return NULL;
860 	}
861 
862 	/* queue up for recycling/reuse */
863 	page_ref_inc(di->page);
864 
865 	skb_reserve(skb, rx_headroom);
866 	skb_put(skb, cqe_bcnt);
867 
868 	return skb;
869 }
870 
871 void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
872 {
873 	struct mlx5e_wqe_frag_info *wi;
874 	struct mlx5e_rx_wqe *wqe;
875 	__be16 wqe_counter_be;
876 	struct sk_buff *skb;
877 	u16 wqe_counter;
878 	u32 cqe_bcnt;
879 
880 	wqe_counter_be = cqe->wqe_counter;
881 	wqe_counter    = be16_to_cpu(wqe_counter_be);
882 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
883 	wi             = &rq->wqe.frag_info[wqe_counter];
884 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
885 
886 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
887 	if (!skb) {
888 		/* probably for XDP */
889 		if (rq->wqe.xdp_xmit) {
890 			wi->di.page = NULL;
891 			rq->wqe.xdp_xmit = false;
892 			/* do not return page to cache, it will be returned on XDP_TX completion */
893 			goto wq_ll_pop;
894 		}
895 		/* probably an XDP_DROP, save the page-reuse checks */
896 		mlx5e_free_rx_wqe(rq, wi);
897 		goto wq_ll_pop;
898 	}
899 
900 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
901 	napi_gro_receive(rq->cq.napi, skb);
902 
903 	mlx5e_free_rx_wqe_reuse(rq, wi);
904 wq_ll_pop:
905 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
906 		       &wqe->next.next_wqe_index);
907 }
908 
909 #ifdef CONFIG_MLX5_ESWITCH
910 void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
911 {
912 	struct net_device *netdev = rq->netdev;
913 	struct mlx5e_priv *priv = netdev_priv(netdev);
914 	struct mlx5e_rep_priv *rpriv  = priv->ppriv;
915 	struct mlx5_eswitch_rep *rep = rpriv->rep;
916 	struct mlx5e_wqe_frag_info *wi;
917 	struct mlx5e_rx_wqe *wqe;
918 	struct sk_buff *skb;
919 	__be16 wqe_counter_be;
920 	u16 wqe_counter;
921 	u32 cqe_bcnt;
922 
923 	wqe_counter_be = cqe->wqe_counter;
924 	wqe_counter    = be16_to_cpu(wqe_counter_be);
925 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
926 	wi             = &rq->wqe.frag_info[wqe_counter];
927 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
928 
929 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
930 	if (!skb) {
931 		if (rq->wqe.xdp_xmit) {
932 			wi->di.page = NULL;
933 			rq->wqe.xdp_xmit = false;
934 			/* do not return page to cache, it will be returned on XDP_TX completion */
935 			goto wq_ll_pop;
936 		}
937 		/* probably an XDP_DROP, save the page-reuse checks */
938 		mlx5e_free_rx_wqe(rq, wi);
939 		goto wq_ll_pop;
940 	}
941 
942 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
943 
944 	if (rep->vlan && skb_vlan_tag_present(skb))
945 		skb_vlan_pop(skb);
946 
947 	napi_gro_receive(rq->cq.napi, skb);
948 
949 	mlx5e_free_rx_wqe_reuse(rq, wi);
950 wq_ll_pop:
951 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
952 		       &wqe->next.next_wqe_index);
953 }
954 #endif
955 
956 static inline void mlx5e_mpwqe_fill_rx_skb(struct mlx5e_rq *rq,
957 					   struct mlx5_cqe64 *cqe,
958 					   struct mlx5e_mpw_info *wi,
959 					   u32 cqe_bcnt,
960 					   struct sk_buff *skb)
961 {
962 	u16 stride_ix      = mpwrq_get_cqe_stride_index(cqe);
963 	u32 wqe_offset     = stride_ix << rq->mpwqe.log_stride_sz;
964 	u32 head_offset    = wqe_offset & (PAGE_SIZE - 1);
965 	u32 page_idx       = wqe_offset >> PAGE_SHIFT;
966 	u32 head_page_idx  = page_idx;
967 	u16 headlen = min_t(u16, MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, cqe_bcnt);
968 	u32 frag_offset    = head_offset + headlen;
969 	u16 byte_cnt       = cqe_bcnt - headlen;
970 
971 	if (unlikely(frag_offset >= PAGE_SIZE)) {
972 		page_idx++;
973 		frag_offset -= PAGE_SIZE;
974 	}
975 
976 	while (byte_cnt) {
977 		u32 pg_consumed_bytes =
978 			min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
979 
980 		mlx5e_add_skb_frag_mpwqe(rq, skb, wi, page_idx, frag_offset,
981 					 pg_consumed_bytes);
982 		byte_cnt -= pg_consumed_bytes;
983 		frag_offset = 0;
984 		page_idx++;
985 	}
986 	/* copy header */
987 	mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, wi, head_page_idx,
988 				    head_offset, headlen);
989 	/* skb linear part was allocated with headlen and aligned to long */
990 	skb->tail += headlen;
991 	skb->len  += headlen;
992 }
993 
994 void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
995 {
996 	u16 cstrides       = mpwrq_get_cqe_consumed_strides(cqe);
997 	u16 wqe_id         = be16_to_cpu(cqe->wqe_id);
998 	struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
999 	struct mlx5e_rx_wqe  *wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_id);
1000 	struct sk_buff *skb;
1001 	u16 cqe_bcnt;
1002 
1003 	wi->consumed_strides += cstrides;
1004 
1005 	if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
1006 		rq->stats.wqe_err++;
1007 		goto mpwrq_cqe_out;
1008 	}
1009 
1010 	if (unlikely(mpwrq_is_filler_cqe(cqe))) {
1011 		rq->stats.mpwqe_filler++;
1012 		goto mpwrq_cqe_out;
1013 	}
1014 
1015 	skb = napi_alloc_skb(rq->cq.napi,
1016 			     ALIGN(MLX5_MPWRQ_SMALL_PACKET_THRESHOLD,
1017 				   sizeof(long)));
1018 	if (unlikely(!skb)) {
1019 		rq->stats.buff_alloc_err++;
1020 		goto mpwrq_cqe_out;
1021 	}
1022 
1023 	prefetchw(skb->data);
1024 	cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
1025 
1026 	mlx5e_mpwqe_fill_rx_skb(rq, cqe, wi, cqe_bcnt, skb);
1027 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1028 	napi_gro_receive(rq->cq.napi, skb);
1029 
1030 mpwrq_cqe_out:
1031 	if (likely(wi->consumed_strides < rq->mpwqe.num_strides))
1032 		return;
1033 
1034 	mlx5e_free_rx_mpwqe(rq, wi);
1035 	mlx5_wq_ll_pop(&rq->wq, cqe->wqe_id, &wqe->next.next_wqe_index);
1036 }
1037 
1038 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
1039 {
1040 	struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
1041 	struct mlx5e_xdpsq *xdpsq;
1042 	struct mlx5_cqe64 *cqe;
1043 	int work_done = 0;
1044 
1045 	if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
1046 		return 0;
1047 
1048 	if (cq->decmprs_left)
1049 		work_done += mlx5e_decompress_cqes_cont(rq, cq, 0, budget);
1050 
1051 	cqe = mlx5_cqwq_get_cqe(&cq->wq);
1052 	if (!cqe)
1053 		return 0;
1054 
1055 	xdpsq = &rq->xdpsq;
1056 
1057 	do {
1058 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
1059 			work_done +=
1060 				mlx5e_decompress_cqes_start(rq, cq,
1061 							    budget - work_done);
1062 			continue;
1063 		}
1064 
1065 		mlx5_cqwq_pop(&cq->wq);
1066 
1067 		rq->handle_rx_cqe(rq, cqe);
1068 	} while ((++work_done < budget) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1069 
1070 	if (xdpsq->db.doorbell) {
1071 		mlx5e_xmit_xdp_doorbell(xdpsq);
1072 		xdpsq->db.doorbell = false;
1073 	}
1074 
1075 	mlx5_cqwq_update_db_record(&cq->wq);
1076 
1077 	/* ensure cq space is freed before enabling more cqes */
1078 	wmb();
1079 
1080 	return work_done;
1081 }
1082 
1083 bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
1084 {
1085 	struct mlx5e_xdpsq *sq;
1086 	struct mlx5_cqe64 *cqe;
1087 	struct mlx5e_rq *rq;
1088 	u16 sqcc;
1089 	int i;
1090 
1091 	sq = container_of(cq, struct mlx5e_xdpsq, cq);
1092 
1093 	if (unlikely(!MLX5E_TEST_BIT(sq->state, MLX5E_SQ_STATE_ENABLED)))
1094 		return false;
1095 
1096 	cqe = mlx5_cqwq_get_cqe(&cq->wq);
1097 	if (!cqe)
1098 		return false;
1099 
1100 	rq = container_of(sq, struct mlx5e_rq, xdpsq);
1101 
1102 	/* sq->cc must be updated only after mlx5_cqwq_update_db_record(),
1103 	 * otherwise a cq overrun may occur
1104 	 */
1105 	sqcc = sq->cc;
1106 
1107 	i = 0;
1108 	do {
1109 		u16 wqe_counter;
1110 		bool last_wqe;
1111 
1112 		mlx5_cqwq_pop(&cq->wq);
1113 
1114 		wqe_counter = be16_to_cpu(cqe->wqe_counter);
1115 
1116 		do {
1117 			struct mlx5e_dma_info *di;
1118 			u16 ci;
1119 
1120 			last_wqe = (sqcc == wqe_counter);
1121 
1122 			ci = sqcc & sq->wq.sz_m1;
1123 			di = &sq->db.di[ci];
1124 
1125 			sqcc++;
1126 			/* Recycle RX page */
1127 			mlx5e_page_release(rq, di, true);
1128 		} while (!last_wqe);
1129 	} while ((++i < MLX5E_TX_CQ_POLL_BUDGET) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1130 
1131 	mlx5_cqwq_update_db_record(&cq->wq);
1132 
1133 	/* ensure cq space is freed before enabling more cqes */
1134 	wmb();
1135 
1136 	sq->cc = sqcc;
1137 	return (i == MLX5E_TX_CQ_POLL_BUDGET);
1138 }
1139 
1140 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
1141 {
1142 	struct mlx5e_rq *rq = container_of(sq, struct mlx5e_rq, xdpsq);
1143 	struct mlx5e_dma_info *di;
1144 	u16 ci;
1145 
1146 	while (sq->cc != sq->pc) {
1147 		ci = sq->cc & sq->wq.sz_m1;
1148 		di = &sq->db.di[ci];
1149 		sq->cc++;
1150 
1151 		mlx5e_page_release(rq, di, false);
1152 	}
1153 }
1154 
1155 #ifdef CONFIG_MLX5_CORE_IPOIB
1156 
1157 #define MLX5_IB_GRH_DGID_OFFSET 24
1158 #define MLX5_GID_SIZE           16
1159 
1160 static inline void mlx5i_complete_rx_cqe(struct mlx5e_rq *rq,
1161 					 struct mlx5_cqe64 *cqe,
1162 					 u32 cqe_bcnt,
1163 					 struct sk_buff *skb)
1164 {
1165 	struct net_device *netdev = rq->netdev;
1166 	struct mlx5e_tstamp *tstamp = rq->tstamp;
1167 	char *pseudo_header;
1168 	u8 *dgid;
1169 	u8 g;
1170 
1171 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
1172 	dgid = skb->data + MLX5_IB_GRH_DGID_OFFSET;
1173 	if ((!g) || dgid[0] != 0xff)
1174 		skb->pkt_type = PACKET_HOST;
1175 	else if (memcmp(dgid, netdev->broadcast + 4, MLX5_GID_SIZE) == 0)
1176 		skb->pkt_type = PACKET_BROADCAST;
1177 	else
1178 		skb->pkt_type = PACKET_MULTICAST;
1179 
1180 	/* TODO: IB/ipoib: Allow mcast packets from other VFs
1181 	 * 68996a6e760e5c74654723eeb57bf65628ae87f4
1182 	 */
1183 
1184 	skb_pull(skb, MLX5_IB_GRH_BYTES);
1185 
1186 	skb->protocol = *((__be16 *)(skb->data));
1187 
1188 	skb->ip_summed = CHECKSUM_COMPLETE;
1189 	skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1190 
1191 	if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
1192 		mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
1193 
1194 	skb_record_rx_queue(skb, rq->ix);
1195 
1196 	if (likely(netdev->features & NETIF_F_RXHASH))
1197 		mlx5e_skb_set_hash(cqe, skb);
1198 
1199 	/* 20 bytes of ipoib header and 4 for encap existing */
1200 	pseudo_header = skb_push(skb, MLX5_IPOIB_PSEUDO_LEN);
1201 	memset(pseudo_header, 0, MLX5_IPOIB_PSEUDO_LEN);
1202 	skb_reset_mac_header(skb);
1203 	skb_pull(skb, MLX5_IPOIB_HARD_LEN);
1204 
1205 	skb->dev = netdev;
1206 
1207 	rq->stats.csum_complete++;
1208 	rq->stats.packets++;
1209 	rq->stats.bytes += cqe_bcnt;
1210 }
1211 
1212 void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1213 {
1214 	struct mlx5e_wqe_frag_info *wi;
1215 	struct mlx5e_rx_wqe *wqe;
1216 	__be16 wqe_counter_be;
1217 	struct sk_buff *skb;
1218 	u16 wqe_counter;
1219 	u32 cqe_bcnt;
1220 
1221 	wqe_counter_be = cqe->wqe_counter;
1222 	wqe_counter    = be16_to_cpu(wqe_counter_be);
1223 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1224 	wi             = &rq->wqe.frag_info[wqe_counter];
1225 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1226 
1227 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1228 	if (!skb)
1229 		goto wq_free_wqe;
1230 
1231 	mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1232 	napi_gro_receive(rq->cq.napi, skb);
1233 
1234 wq_free_wqe:
1235 	mlx5e_free_rx_wqe_reuse(rq, wi);
1236 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1237 		       &wqe->next.next_wqe_index);
1238 }
1239 
1240 #endif /* CONFIG_MLX5_CORE_IPOIB */
1241 
1242 #ifdef CONFIG_MLX5_EN_IPSEC
1243 
1244 void mlx5e_ipsec_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1245 {
1246 	struct mlx5e_wqe_frag_info *wi;
1247 	struct mlx5e_rx_wqe *wqe;
1248 	__be16 wqe_counter_be;
1249 	struct sk_buff *skb;
1250 	u16 wqe_counter;
1251 	u32 cqe_bcnt;
1252 
1253 	wqe_counter_be = cqe->wqe_counter;
1254 	wqe_counter    = be16_to_cpu(wqe_counter_be);
1255 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1256 	wi             = &rq->wqe.frag_info[wqe_counter];
1257 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1258 
1259 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1260 	if (unlikely(!skb)) {
1261 		/* a DROP, save the page-reuse checks */
1262 		mlx5e_free_rx_wqe(rq, wi);
1263 		goto wq_ll_pop;
1264 	}
1265 	skb = mlx5e_ipsec_handle_rx_skb(rq->netdev, skb);
1266 	if (unlikely(!skb)) {
1267 		mlx5e_free_rx_wqe(rq, wi);
1268 		goto wq_ll_pop;
1269 	}
1270 
1271 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1272 	napi_gro_receive(rq->cq.napi, skb);
1273 
1274 	mlx5e_free_rx_wqe_reuse(rq, wi);
1275 wq_ll_pop:
1276 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1277 		       &wqe->next.next_wqe_index);
1278 }
1279 
1280 #endif /* CONFIG_MLX5_EN_IPSEC */
1281