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 		return;
631 	}
632 
633 	if (is_first_ethertype_ip(skb)) {
634 		skb->ip_summed = CHECKSUM_COMPLETE;
635 		skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
636 		rq->stats.csum_complete++;
637 		return;
638 	}
639 
640 	if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
641 		   (cqe->hds_ip_ext & CQE_L4_OK))) {
642 		skb->ip_summed = CHECKSUM_UNNECESSARY;
643 		if (cqe_is_tunneled(cqe)) {
644 			skb->csum_level = 1;
645 			skb->encapsulation = 1;
646 			rq->stats.csum_unnecessary_inner++;
647 		}
648 		return;
649 	}
650 csum_none:
651 	skb->ip_summed = CHECKSUM_NONE;
652 	rq->stats.csum_none++;
653 }
654 
655 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
656 				      u32 cqe_bcnt,
657 				      struct mlx5e_rq *rq,
658 				      struct sk_buff *skb)
659 {
660 	struct net_device *netdev = rq->netdev;
661 	struct mlx5e_tstamp *tstamp = rq->tstamp;
662 	int lro_num_seg;
663 
664 	lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
665 	if (lro_num_seg > 1) {
666 		mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
667 		skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
668 		/* Subtract one since we already counted this as one
669 		 * "regular" packet in mlx5e_complete_rx_cqe()
670 		 */
671 		rq->stats.packets += lro_num_seg - 1;
672 		rq->stats.lro_packets++;
673 		rq->stats.lro_bytes += cqe_bcnt;
674 	}
675 
676 	if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
677 		mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
678 
679 	skb_record_rx_queue(skb, rq->ix);
680 
681 	if (likely(netdev->features & NETIF_F_RXHASH))
682 		mlx5e_skb_set_hash(cqe, skb);
683 
684 	if (cqe_has_vlan(cqe))
685 		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
686 				       be16_to_cpu(cqe->vlan_info));
687 
688 	skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
689 
690 	mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
691 	skb->protocol = eth_type_trans(skb, netdev);
692 }
693 
694 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
695 					 struct mlx5_cqe64 *cqe,
696 					 u32 cqe_bcnt,
697 					 struct sk_buff *skb)
698 {
699 	rq->stats.packets++;
700 	rq->stats.bytes += cqe_bcnt;
701 	mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
702 }
703 
704 static inline void mlx5e_xmit_xdp_doorbell(struct mlx5e_xdpsq *sq)
705 {
706 	struct mlx5_wq_cyc *wq = &sq->wq;
707 	struct mlx5e_tx_wqe *wqe;
708 	u16 pi = (sq->pc - 1) & wq->sz_m1; /* last pi */
709 
710 	wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
711 
712 	mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &wqe->ctrl);
713 }
714 
715 static inline bool mlx5e_xmit_xdp_frame(struct mlx5e_rq *rq,
716 					struct mlx5e_dma_info *di,
717 					const struct xdp_buff *xdp)
718 {
719 	struct mlx5e_xdpsq       *sq   = &rq->xdpsq;
720 	struct mlx5_wq_cyc       *wq   = &sq->wq;
721 	u16                       pi   = sq->pc & wq->sz_m1;
722 	struct mlx5e_tx_wqe      *wqe  = mlx5_wq_cyc_get_wqe(wq, pi);
723 
724 	struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
725 	struct mlx5_wqe_eth_seg  *eseg = &wqe->eth;
726 	struct mlx5_wqe_data_seg *dseg;
727 
728 	ptrdiff_t data_offset = xdp->data - xdp->data_hard_start;
729 	dma_addr_t dma_addr  = di->addr + data_offset;
730 	unsigned int dma_len = xdp->data_end - xdp->data;
731 
732 	prefetchw(wqe);
733 
734 	if (unlikely(dma_len < MLX5E_XDP_MIN_INLINE ||
735 		     MLX5E_SW2HW_MTU(rq->channel->priv, rq->netdev->mtu) < dma_len)) {
736 		rq->stats.xdp_drop++;
737 		return false;
738 	}
739 
740 	if (unlikely(!mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1))) {
741 		if (sq->db.doorbell) {
742 			/* SQ is full, ring doorbell */
743 			mlx5e_xmit_xdp_doorbell(sq);
744 			sq->db.doorbell = false;
745 		}
746 		rq->stats.xdp_tx_full++;
747 		return false;
748 	}
749 
750 	dma_sync_single_for_device(sq->pdev, dma_addr, dma_len, PCI_DMA_TODEVICE);
751 
752 	cseg->fm_ce_se = 0;
753 
754 	dseg = (struct mlx5_wqe_data_seg *)eseg + 1;
755 
756 	/* copy the inline part if required */
757 	if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
758 		memcpy(eseg->inline_hdr.start, xdp->data, MLX5E_XDP_MIN_INLINE);
759 		eseg->inline_hdr.sz = cpu_to_be16(MLX5E_XDP_MIN_INLINE);
760 		dma_len  -= MLX5E_XDP_MIN_INLINE;
761 		dma_addr += MLX5E_XDP_MIN_INLINE;
762 		dseg++;
763 	}
764 
765 	/* write the dma part */
766 	dseg->addr       = cpu_to_be64(dma_addr);
767 	dseg->byte_count = cpu_to_be32(dma_len);
768 
769 	cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_SEND);
770 
771 	/* move page to reference to sq responsibility,
772 	 * and mark so it's not put back in page-cache.
773 	 */
774 	rq->wqe.xdp_xmit = true;
775 	sq->db.di[pi] = *di;
776 	sq->pc++;
777 
778 	sq->db.doorbell = true;
779 
780 	rq->stats.xdp_tx++;
781 	return true;
782 }
783 
784 /* returns true if packet was consumed by xdp */
785 static inline int mlx5e_xdp_handle(struct mlx5e_rq *rq,
786 				   struct mlx5e_dma_info *di,
787 				   void *va, u16 *rx_headroom, u32 *len)
788 {
789 	const struct bpf_prog *prog = READ_ONCE(rq->xdp_prog);
790 	struct xdp_buff xdp;
791 	u32 act;
792 
793 	if (!prog)
794 		return false;
795 
796 	xdp.data = va + *rx_headroom;
797 	xdp_set_data_meta_invalid(&xdp);
798 	xdp.data_end = xdp.data + *len;
799 	xdp.data_hard_start = va;
800 
801 	act = bpf_prog_run_xdp(prog, &xdp);
802 	switch (act) {
803 	case XDP_PASS:
804 		*rx_headroom = xdp.data - xdp.data_hard_start;
805 		*len = xdp.data_end - xdp.data;
806 		return false;
807 	case XDP_TX:
808 		if (unlikely(!mlx5e_xmit_xdp_frame(rq, di, &xdp)))
809 			trace_xdp_exception(rq->netdev, prog, act);
810 		return true;
811 	default:
812 		bpf_warn_invalid_xdp_action(act);
813 	case XDP_ABORTED:
814 		trace_xdp_exception(rq->netdev, prog, act);
815 	case XDP_DROP:
816 		rq->stats.xdp_drop++;
817 		return true;
818 	}
819 }
820 
821 static inline
822 struct sk_buff *skb_from_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
823 			     struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
824 {
825 	struct mlx5e_dma_info *di = &wi->di;
826 	u16 rx_headroom = rq->buff.headroom;
827 	struct sk_buff *skb;
828 	void *va, *data;
829 	bool consumed;
830 	u32 frag_size;
831 
832 	va             = page_address(di->page) + wi->offset;
833 	data           = va + rx_headroom;
834 	frag_size      = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
835 
836 	dma_sync_single_range_for_cpu(rq->pdev,
837 				      di->addr + wi->offset,
838 				      0, frag_size,
839 				      DMA_FROM_DEVICE);
840 	prefetch(data);
841 	wi->offset += frag_size;
842 
843 	if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
844 		rq->stats.wqe_err++;
845 		return NULL;
846 	}
847 
848 	rcu_read_lock();
849 	consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt);
850 	rcu_read_unlock();
851 	if (consumed)
852 		return NULL; /* page/packet was consumed by XDP */
853 
854 	skb = build_skb(va, frag_size);
855 	if (unlikely(!skb)) {
856 		rq->stats.buff_alloc_err++;
857 		return NULL;
858 	}
859 
860 	/* queue up for recycling/reuse */
861 	page_ref_inc(di->page);
862 
863 	skb_reserve(skb, rx_headroom);
864 	skb_put(skb, cqe_bcnt);
865 
866 	return skb;
867 }
868 
869 void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
870 {
871 	struct mlx5e_wqe_frag_info *wi;
872 	struct mlx5e_rx_wqe *wqe;
873 	__be16 wqe_counter_be;
874 	struct sk_buff *skb;
875 	u16 wqe_counter;
876 	u32 cqe_bcnt;
877 
878 	wqe_counter_be = cqe->wqe_counter;
879 	wqe_counter    = be16_to_cpu(wqe_counter_be);
880 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
881 	wi             = &rq->wqe.frag_info[wqe_counter];
882 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
883 
884 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
885 	if (!skb) {
886 		/* probably for XDP */
887 		if (rq->wqe.xdp_xmit) {
888 			wi->di.page = NULL;
889 			rq->wqe.xdp_xmit = false;
890 			/* do not return page to cache, it will be returned on XDP_TX completion */
891 			goto wq_ll_pop;
892 		}
893 		/* probably an XDP_DROP, save the page-reuse checks */
894 		mlx5e_free_rx_wqe(rq, wi);
895 		goto wq_ll_pop;
896 	}
897 
898 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
899 	napi_gro_receive(rq->cq.napi, skb);
900 
901 	mlx5e_free_rx_wqe_reuse(rq, wi);
902 wq_ll_pop:
903 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
904 		       &wqe->next.next_wqe_index);
905 }
906 
907 #ifdef CONFIG_MLX5_ESWITCH
908 void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
909 {
910 	struct net_device *netdev = rq->netdev;
911 	struct mlx5e_priv *priv = netdev_priv(netdev);
912 	struct mlx5e_rep_priv *rpriv  = priv->ppriv;
913 	struct mlx5_eswitch_rep *rep = rpriv->rep;
914 	struct mlx5e_wqe_frag_info *wi;
915 	struct mlx5e_rx_wqe *wqe;
916 	struct sk_buff *skb;
917 	__be16 wqe_counter_be;
918 	u16 wqe_counter;
919 	u32 cqe_bcnt;
920 
921 	wqe_counter_be = cqe->wqe_counter;
922 	wqe_counter    = be16_to_cpu(wqe_counter_be);
923 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
924 	wi             = &rq->wqe.frag_info[wqe_counter];
925 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
926 
927 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
928 	if (!skb) {
929 		if (rq->wqe.xdp_xmit) {
930 			wi->di.page = NULL;
931 			rq->wqe.xdp_xmit = false;
932 			/* do not return page to cache, it will be returned on XDP_TX completion */
933 			goto wq_ll_pop;
934 		}
935 		/* probably an XDP_DROP, save the page-reuse checks */
936 		mlx5e_free_rx_wqe(rq, wi);
937 		goto wq_ll_pop;
938 	}
939 
940 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
941 
942 	if (rep->vlan && skb_vlan_tag_present(skb))
943 		skb_vlan_pop(skb);
944 
945 	napi_gro_receive(rq->cq.napi, skb);
946 
947 	mlx5e_free_rx_wqe_reuse(rq, wi);
948 wq_ll_pop:
949 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
950 		       &wqe->next.next_wqe_index);
951 }
952 #endif
953 
954 static inline void mlx5e_mpwqe_fill_rx_skb(struct mlx5e_rq *rq,
955 					   struct mlx5_cqe64 *cqe,
956 					   struct mlx5e_mpw_info *wi,
957 					   u32 cqe_bcnt,
958 					   struct sk_buff *skb)
959 {
960 	u16 stride_ix      = mpwrq_get_cqe_stride_index(cqe);
961 	u32 wqe_offset     = stride_ix << rq->mpwqe.log_stride_sz;
962 	u32 head_offset    = wqe_offset & (PAGE_SIZE - 1);
963 	u32 page_idx       = wqe_offset >> PAGE_SHIFT;
964 	u32 head_page_idx  = page_idx;
965 	u16 headlen = min_t(u16, MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, cqe_bcnt);
966 	u32 frag_offset    = head_offset + headlen;
967 	u16 byte_cnt       = cqe_bcnt - headlen;
968 
969 	if (unlikely(frag_offset >= PAGE_SIZE)) {
970 		page_idx++;
971 		frag_offset -= PAGE_SIZE;
972 	}
973 
974 	while (byte_cnt) {
975 		u32 pg_consumed_bytes =
976 			min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
977 
978 		mlx5e_add_skb_frag_mpwqe(rq, skb, wi, page_idx, frag_offset,
979 					 pg_consumed_bytes);
980 		byte_cnt -= pg_consumed_bytes;
981 		frag_offset = 0;
982 		page_idx++;
983 	}
984 	/* copy header */
985 	mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, wi, head_page_idx,
986 				    head_offset, headlen);
987 	/* skb linear part was allocated with headlen and aligned to long */
988 	skb->tail += headlen;
989 	skb->len  += headlen;
990 }
991 
992 void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
993 {
994 	u16 cstrides       = mpwrq_get_cqe_consumed_strides(cqe);
995 	u16 wqe_id         = be16_to_cpu(cqe->wqe_id);
996 	struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
997 	struct mlx5e_rx_wqe  *wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_id);
998 	struct sk_buff *skb;
999 	u16 cqe_bcnt;
1000 
1001 	wi->consumed_strides += cstrides;
1002 
1003 	if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
1004 		rq->stats.wqe_err++;
1005 		goto mpwrq_cqe_out;
1006 	}
1007 
1008 	if (unlikely(mpwrq_is_filler_cqe(cqe))) {
1009 		rq->stats.mpwqe_filler++;
1010 		goto mpwrq_cqe_out;
1011 	}
1012 
1013 	skb = napi_alloc_skb(rq->cq.napi,
1014 			     ALIGN(MLX5_MPWRQ_SMALL_PACKET_THRESHOLD,
1015 				   sizeof(long)));
1016 	if (unlikely(!skb)) {
1017 		rq->stats.buff_alloc_err++;
1018 		goto mpwrq_cqe_out;
1019 	}
1020 
1021 	prefetchw(skb->data);
1022 	cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
1023 
1024 	mlx5e_mpwqe_fill_rx_skb(rq, cqe, wi, cqe_bcnt, skb);
1025 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1026 	napi_gro_receive(rq->cq.napi, skb);
1027 
1028 mpwrq_cqe_out:
1029 	if (likely(wi->consumed_strides < rq->mpwqe.num_strides))
1030 		return;
1031 
1032 	mlx5e_free_rx_mpwqe(rq, wi);
1033 	mlx5_wq_ll_pop(&rq->wq, cqe->wqe_id, &wqe->next.next_wqe_index);
1034 }
1035 
1036 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
1037 {
1038 	struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
1039 	struct mlx5e_xdpsq *xdpsq;
1040 	struct mlx5_cqe64 *cqe;
1041 	int work_done = 0;
1042 
1043 	if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
1044 		return 0;
1045 
1046 	if (cq->decmprs_left)
1047 		work_done += mlx5e_decompress_cqes_cont(rq, cq, 0, budget);
1048 
1049 	cqe = mlx5_cqwq_get_cqe(&cq->wq);
1050 	if (!cqe)
1051 		return 0;
1052 
1053 	xdpsq = &rq->xdpsq;
1054 
1055 	do {
1056 		if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
1057 			work_done +=
1058 				mlx5e_decompress_cqes_start(rq, cq,
1059 							    budget - work_done);
1060 			continue;
1061 		}
1062 
1063 		mlx5_cqwq_pop(&cq->wq);
1064 
1065 		rq->handle_rx_cqe(rq, cqe);
1066 	} while ((++work_done < budget) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1067 
1068 	if (xdpsq->db.doorbell) {
1069 		mlx5e_xmit_xdp_doorbell(xdpsq);
1070 		xdpsq->db.doorbell = false;
1071 	}
1072 
1073 	mlx5_cqwq_update_db_record(&cq->wq);
1074 
1075 	/* ensure cq space is freed before enabling more cqes */
1076 	wmb();
1077 
1078 	return work_done;
1079 }
1080 
1081 bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
1082 {
1083 	struct mlx5e_xdpsq *sq;
1084 	struct mlx5_cqe64 *cqe;
1085 	struct mlx5e_rq *rq;
1086 	u16 sqcc;
1087 	int i;
1088 
1089 	sq = container_of(cq, struct mlx5e_xdpsq, cq);
1090 
1091 	if (unlikely(!MLX5E_TEST_BIT(sq->state, MLX5E_SQ_STATE_ENABLED)))
1092 		return false;
1093 
1094 	cqe = mlx5_cqwq_get_cqe(&cq->wq);
1095 	if (!cqe)
1096 		return false;
1097 
1098 	rq = container_of(sq, struct mlx5e_rq, xdpsq);
1099 
1100 	/* sq->cc must be updated only after mlx5_cqwq_update_db_record(),
1101 	 * otherwise a cq overrun may occur
1102 	 */
1103 	sqcc = sq->cc;
1104 
1105 	i = 0;
1106 	do {
1107 		u16 wqe_counter;
1108 		bool last_wqe;
1109 
1110 		mlx5_cqwq_pop(&cq->wq);
1111 
1112 		wqe_counter = be16_to_cpu(cqe->wqe_counter);
1113 
1114 		do {
1115 			struct mlx5e_dma_info *di;
1116 			u16 ci;
1117 
1118 			last_wqe = (sqcc == wqe_counter);
1119 
1120 			ci = sqcc & sq->wq.sz_m1;
1121 			di = &sq->db.di[ci];
1122 
1123 			sqcc++;
1124 			/* Recycle RX page */
1125 			mlx5e_page_release(rq, di, true);
1126 		} while (!last_wqe);
1127 	} while ((++i < MLX5E_TX_CQ_POLL_BUDGET) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1128 
1129 	mlx5_cqwq_update_db_record(&cq->wq);
1130 
1131 	/* ensure cq space is freed before enabling more cqes */
1132 	wmb();
1133 
1134 	sq->cc = sqcc;
1135 	return (i == MLX5E_TX_CQ_POLL_BUDGET);
1136 }
1137 
1138 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
1139 {
1140 	struct mlx5e_rq *rq = container_of(sq, struct mlx5e_rq, xdpsq);
1141 	struct mlx5e_dma_info *di;
1142 	u16 ci;
1143 
1144 	while (sq->cc != sq->pc) {
1145 		ci = sq->cc & sq->wq.sz_m1;
1146 		di = &sq->db.di[ci];
1147 		sq->cc++;
1148 
1149 		mlx5e_page_release(rq, di, false);
1150 	}
1151 }
1152 
1153 #ifdef CONFIG_MLX5_CORE_IPOIB
1154 
1155 #define MLX5_IB_GRH_DGID_OFFSET 24
1156 #define MLX5_GID_SIZE           16
1157 
1158 static inline void mlx5i_complete_rx_cqe(struct mlx5e_rq *rq,
1159 					 struct mlx5_cqe64 *cqe,
1160 					 u32 cqe_bcnt,
1161 					 struct sk_buff *skb)
1162 {
1163 	struct net_device *netdev = rq->netdev;
1164 	struct mlx5e_tstamp *tstamp = rq->tstamp;
1165 	char *pseudo_header;
1166 	u8 *dgid;
1167 	u8 g;
1168 
1169 	g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
1170 	dgid = skb->data + MLX5_IB_GRH_DGID_OFFSET;
1171 	if ((!g) || dgid[0] != 0xff)
1172 		skb->pkt_type = PACKET_HOST;
1173 	else if (memcmp(dgid, netdev->broadcast + 4, MLX5_GID_SIZE) == 0)
1174 		skb->pkt_type = PACKET_BROADCAST;
1175 	else
1176 		skb->pkt_type = PACKET_MULTICAST;
1177 
1178 	/* TODO: IB/ipoib: Allow mcast packets from other VFs
1179 	 * 68996a6e760e5c74654723eeb57bf65628ae87f4
1180 	 */
1181 
1182 	skb_pull(skb, MLX5_IB_GRH_BYTES);
1183 
1184 	skb->protocol = *((__be16 *)(skb->data));
1185 
1186 	skb->ip_summed = CHECKSUM_COMPLETE;
1187 	skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1188 
1189 	if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
1190 		mlx5e_fill_hwstamp(tstamp, get_cqe_ts(cqe), skb_hwtstamps(skb));
1191 
1192 	skb_record_rx_queue(skb, rq->ix);
1193 
1194 	if (likely(netdev->features & NETIF_F_RXHASH))
1195 		mlx5e_skb_set_hash(cqe, skb);
1196 
1197 	/* 20 bytes of ipoib header and 4 for encap existing */
1198 	pseudo_header = skb_push(skb, MLX5_IPOIB_PSEUDO_LEN);
1199 	memset(pseudo_header, 0, MLX5_IPOIB_PSEUDO_LEN);
1200 	skb_reset_mac_header(skb);
1201 	skb_pull(skb, MLX5_IPOIB_HARD_LEN);
1202 
1203 	skb->dev = netdev;
1204 
1205 	rq->stats.csum_complete++;
1206 	rq->stats.packets++;
1207 	rq->stats.bytes += cqe_bcnt;
1208 }
1209 
1210 void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1211 {
1212 	struct mlx5e_wqe_frag_info *wi;
1213 	struct mlx5e_rx_wqe *wqe;
1214 	__be16 wqe_counter_be;
1215 	struct sk_buff *skb;
1216 	u16 wqe_counter;
1217 	u32 cqe_bcnt;
1218 
1219 	wqe_counter_be = cqe->wqe_counter;
1220 	wqe_counter    = be16_to_cpu(wqe_counter_be);
1221 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1222 	wi             = &rq->wqe.frag_info[wqe_counter];
1223 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1224 
1225 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1226 	if (!skb)
1227 		goto wq_free_wqe;
1228 
1229 	mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1230 	napi_gro_receive(rq->cq.napi, skb);
1231 
1232 wq_free_wqe:
1233 	mlx5e_free_rx_wqe_reuse(rq, wi);
1234 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1235 		       &wqe->next.next_wqe_index);
1236 }
1237 
1238 #endif /* CONFIG_MLX5_CORE_IPOIB */
1239 
1240 #ifdef CONFIG_MLX5_EN_IPSEC
1241 
1242 void mlx5e_ipsec_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1243 {
1244 	struct mlx5e_wqe_frag_info *wi;
1245 	struct mlx5e_rx_wqe *wqe;
1246 	__be16 wqe_counter_be;
1247 	struct sk_buff *skb;
1248 	u16 wqe_counter;
1249 	u32 cqe_bcnt;
1250 
1251 	wqe_counter_be = cqe->wqe_counter;
1252 	wqe_counter    = be16_to_cpu(wqe_counter_be);
1253 	wqe            = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1254 	wi             = &rq->wqe.frag_info[wqe_counter];
1255 	cqe_bcnt       = be32_to_cpu(cqe->byte_cnt);
1256 
1257 	skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1258 	if (unlikely(!skb)) {
1259 		/* a DROP, save the page-reuse checks */
1260 		mlx5e_free_rx_wqe(rq, wi);
1261 		goto wq_ll_pop;
1262 	}
1263 	skb = mlx5e_ipsec_handle_rx_skb(rq->netdev, skb);
1264 	if (unlikely(!skb)) {
1265 		mlx5e_free_rx_wqe(rq, wi);
1266 		goto wq_ll_pop;
1267 	}
1268 
1269 	mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1270 	napi_gro_receive(rq->cq.napi, skb);
1271 
1272 	mlx5e_free_rx_wqe_reuse(rq, wi);
1273 wq_ll_pop:
1274 	mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1275 		       &wqe->next.next_wqe_index);
1276 }
1277 
1278 #endif /* CONFIG_MLX5_EN_IPSEC */
1279