2 * Copyright (c) 2015, Mellanox Technologies. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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.
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
33 #include <linux/prefetch.h>
35 #include <linux/ipv6.h>
36 #include <linux/tcp.h>
37 #include <net/ip6_checksum.h>
38 #include <net/page_pool.h>
39 #include <net/inet_ecn.h>
44 #include "ipoib/ipoib.h"
45 #include "en_accel/ipsec_rxtx.h"
46 #include "en_accel/tls_rxtx.h"
47 #include "lib/clock.h"
50 static inline bool mlx5e_rx_hw_stamp(struct hwtstamp_config *config)
52 return config->rx_filter == HWTSTAMP_FILTER_ALL;
55 static inline void mlx5e_read_cqe_slot(struct mlx5_cqwq *wq,
58 u32 ci = mlx5_cqwq_ctr2ix(wq, cqcc);
60 memcpy(data, mlx5_cqwq_get_wqe(wq, ci), sizeof(struct mlx5_cqe64));
63 static inline void mlx5e_read_title_slot(struct mlx5e_rq *rq,
67 struct mlx5e_cq_decomp *cqd = &rq->cqd;
68 struct mlx5_cqe64 *title = &cqd->title;
70 mlx5e_read_cqe_slot(wq, cqcc, title);
71 cqd->left = be32_to_cpu(title->byte_cnt);
72 cqd->wqe_counter = be16_to_cpu(title->wqe_counter);
73 rq->stats->cqe_compress_blks++;
76 static inline void mlx5e_read_mini_arr_slot(struct mlx5_cqwq *wq,
77 struct mlx5e_cq_decomp *cqd,
80 mlx5e_read_cqe_slot(wq, cqcc, cqd->mini_arr);
81 cqd->mini_arr_idx = 0;
84 static inline void mlx5e_cqes_update_owner(struct mlx5_cqwq *wq, int n)
87 u8 op_own = mlx5_cqwq_get_ctr_wrap_cnt(wq, cqcc) & 1;
88 u32 ci = mlx5_cqwq_ctr2ix(wq, cqcc);
89 u32 wq_sz = mlx5_cqwq_get_size(wq);
90 u32 ci_top = min_t(u32, wq_sz, ci + n);
92 for (; ci < ci_top; ci++, n--) {
93 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(wq, ci);
98 if (unlikely(ci == wq_sz)) {
100 for (ci = 0; ci < n; ci++) {
101 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(wq, ci);
103 cqe->op_own = op_own;
108 static inline void mlx5e_decompress_cqe(struct mlx5e_rq *rq,
109 struct mlx5_cqwq *wq,
112 struct mlx5e_cq_decomp *cqd = &rq->cqd;
113 struct mlx5_mini_cqe8 *mini_cqe = &cqd->mini_arr[cqd->mini_arr_idx];
114 struct mlx5_cqe64 *title = &cqd->title;
116 title->byte_cnt = mini_cqe->byte_cnt;
117 title->check_sum = mini_cqe->checksum;
118 title->op_own &= 0xf0;
119 title->op_own |= 0x01 & (cqcc >> wq->fbc.log_sz);
120 title->wqe_counter = cpu_to_be16(cqd->wqe_counter);
122 if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
123 cqd->wqe_counter += mpwrq_get_cqe_consumed_strides(title);
126 mlx5_wq_cyc_ctr2ix(&rq->wqe.wq, cqd->wqe_counter + 1);
129 static inline void mlx5e_decompress_cqe_no_hash(struct mlx5e_rq *rq,
130 struct mlx5_cqwq *wq,
133 struct mlx5e_cq_decomp *cqd = &rq->cqd;
135 mlx5e_decompress_cqe(rq, wq, cqcc);
136 cqd->title.rss_hash_type = 0;
137 cqd->title.rss_hash_result = 0;
140 static inline u32 mlx5e_decompress_cqes_cont(struct mlx5e_rq *rq,
141 struct mlx5_cqwq *wq,
142 int update_owner_only,
145 struct mlx5e_cq_decomp *cqd = &rq->cqd;
146 u32 cqcc = wq->cc + update_owner_only;
150 cqe_count = min_t(u32, cqd->left, budget_rem);
152 for (i = update_owner_only; i < cqe_count;
153 i++, cqd->mini_arr_idx++, cqcc++) {
154 if (cqd->mini_arr_idx == MLX5_MINI_CQE_ARRAY_SIZE)
155 mlx5e_read_mini_arr_slot(wq, cqd, cqcc);
157 mlx5e_decompress_cqe_no_hash(rq, wq, cqcc);
158 rq->handle_rx_cqe(rq, &cqd->title);
160 mlx5e_cqes_update_owner(wq, cqcc - wq->cc);
162 cqd->left -= cqe_count;
163 rq->stats->cqe_compress_pkts += cqe_count;
168 static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
169 struct mlx5_cqwq *wq,
172 struct mlx5e_cq_decomp *cqd = &rq->cqd;
175 mlx5e_read_title_slot(rq, wq, cc);
176 mlx5e_read_mini_arr_slot(wq, cqd, cc + 1);
177 mlx5e_decompress_cqe(rq, wq, cc);
178 rq->handle_rx_cqe(rq, &cqd->title);
181 return mlx5e_decompress_cqes_cont(rq, wq, 1, budget_rem) - 1;
184 static inline bool mlx5e_page_is_reserved(struct page *page)
186 return page_is_pfmemalloc(page) || page_to_nid(page) != numa_mem_id();
189 static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq,
190 struct mlx5e_dma_info *dma_info)
192 struct mlx5e_page_cache *cache = &rq->page_cache;
193 u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
194 struct mlx5e_rq_stats *stats = rq->stats;
196 if (tail_next == cache->head) {
201 if (unlikely(mlx5e_page_is_reserved(dma_info->page))) {
202 stats->cache_waive++;
206 cache->page_cache[cache->tail] = *dma_info;
207 cache->tail = tail_next;
211 static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq,
212 struct mlx5e_dma_info *dma_info)
214 struct mlx5e_page_cache *cache = &rq->page_cache;
215 struct mlx5e_rq_stats *stats = rq->stats;
217 if (unlikely(cache->head == cache->tail)) {
218 stats->cache_empty++;
222 if (page_ref_count(cache->page_cache[cache->head].page) != 1) {
227 *dma_info = cache->page_cache[cache->head];
228 cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
229 stats->cache_reuse++;
231 dma_sync_single_for_device(rq->pdev, dma_info->addr,
237 static inline int mlx5e_page_alloc_mapped(struct mlx5e_rq *rq,
238 struct mlx5e_dma_info *dma_info)
240 if (mlx5e_rx_cache_get(rq, dma_info))
243 dma_info->page = page_pool_dev_alloc_pages(rq->page_pool);
244 if (unlikely(!dma_info->page))
247 dma_info->addr = dma_map_page(rq->pdev, dma_info->page, 0,
248 PAGE_SIZE, rq->buff.map_dir);
249 if (unlikely(dma_mapping_error(rq->pdev, dma_info->addr))) {
250 put_page(dma_info->page);
251 dma_info->page = NULL;
258 void mlx5e_page_dma_unmap(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info)
260 dma_unmap_page(rq->pdev, dma_info->addr, PAGE_SIZE, rq->buff.map_dir);
263 void mlx5e_page_release(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info,
266 if (likely(recycle)) {
267 if (mlx5e_rx_cache_put(rq, dma_info))
270 mlx5e_page_dma_unmap(rq, dma_info);
271 page_pool_recycle_direct(rq->page_pool, dma_info->page);
273 mlx5e_page_dma_unmap(rq, dma_info);
274 put_page(dma_info->page);
278 static inline int mlx5e_get_rx_frag(struct mlx5e_rq *rq,
279 struct mlx5e_wqe_frag_info *frag)
284 /* On first frag (offset == 0), replenish page (dma_info actually).
285 * Other frags that point to the same dma_info (with a different
286 * offset) should just use the new one without replenishing again
289 err = mlx5e_page_alloc_mapped(rq, frag->di);
294 static inline void mlx5e_put_rx_frag(struct mlx5e_rq *rq,
295 struct mlx5e_wqe_frag_info *frag,
298 if (frag->last_in_page)
299 mlx5e_page_release(rq, frag->di, recycle);
302 static inline struct mlx5e_wqe_frag_info *get_frag(struct mlx5e_rq *rq, u16 ix)
304 return &rq->wqe.frags[ix << rq->wqe.info.log_num_frags];
307 static int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe_cyc *wqe,
310 struct mlx5e_wqe_frag_info *frag = get_frag(rq, ix);
314 for (i = 0; i < rq->wqe.info.num_frags; i++, frag++) {
315 err = mlx5e_get_rx_frag(rq, frag);
319 wqe->data[i].addr = cpu_to_be64(frag->di->addr +
320 frag->offset + rq->buff.headroom);
327 mlx5e_put_rx_frag(rq, --frag, true);
332 static inline void mlx5e_free_rx_wqe(struct mlx5e_rq *rq,
333 struct mlx5e_wqe_frag_info *wi,
338 for (i = 0; i < rq->wqe.info.num_frags; i++, wi++)
339 mlx5e_put_rx_frag(rq, wi, recycle);
342 void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
344 struct mlx5e_wqe_frag_info *wi = get_frag(rq, ix);
346 mlx5e_free_rx_wqe(rq, wi, false);
349 static int mlx5e_alloc_rx_wqes(struct mlx5e_rq *rq, u16 ix, u8 wqe_bulk)
351 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
355 for (i = 0; i < wqe_bulk; i++) {
356 struct mlx5e_rx_wqe_cyc *wqe = mlx5_wq_cyc_get_wqe(wq, ix + i);
358 err = mlx5e_alloc_rx_wqe(rq, wqe, ix + i);
367 mlx5e_dealloc_rx_wqe(rq, ix + i);
373 mlx5e_add_skb_frag(struct mlx5e_rq *rq, struct sk_buff *skb,
374 struct mlx5e_dma_info *di, u32 frag_offset, u32 len,
375 unsigned int truesize)
377 dma_sync_single_for_cpu(rq->pdev,
378 di->addr + frag_offset,
379 len, DMA_FROM_DEVICE);
380 page_ref_inc(di->page);
381 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
382 di->page, frag_offset, len, truesize);
386 mlx5e_copy_skb_header(struct device *pdev, struct sk_buff *skb,
387 struct mlx5e_dma_info *dma_info,
388 int offset_from, u32 headlen)
390 const void *from = page_address(dma_info->page) + offset_from;
391 /* Aligning len to sizeof(long) optimizes memcpy performance */
392 unsigned int len = ALIGN(headlen, sizeof(long));
394 dma_sync_single_for_cpu(pdev, dma_info->addr + offset_from, len,
396 skb_copy_to_linear_data(skb, from, len);
400 mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi, bool recycle)
402 const bool no_xdp_xmit =
403 bitmap_empty(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
404 struct mlx5e_dma_info *dma_info = wi->umr.dma_info;
407 for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++)
408 if (no_xdp_xmit || !test_bit(i, wi->xdp_xmit_bitmap))
409 mlx5e_page_release(rq, &dma_info[i], recycle);
412 static void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
414 struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
415 struct mlx5e_rx_wqe_ll *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
417 rq->mpwqe.umr_in_progress = false;
419 mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
421 /* ensure wqes are visible to device before updating doorbell record */
424 mlx5_wq_ll_update_db_record(wq);
427 static inline u16 mlx5e_icosq_wrap_cnt(struct mlx5e_icosq *sq)
429 return sq->pc >> MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
432 static inline void mlx5e_fill_icosq_frag_edge(struct mlx5e_icosq *sq,
433 struct mlx5_wq_cyc *wq,
436 struct mlx5e_sq_wqe_info *edge_wi, *wi = &sq->db.ico_wqe[pi];
438 edge_wi = wi + nnops;
440 /* fill sq frag edge with nops to avoid wqe wrapping two pages */
441 for (; wi < edge_wi; wi++) {
442 wi->opcode = MLX5_OPCODE_NOP;
443 mlx5e_post_nop(wq, sq->sqn, &sq->pc);
447 static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
449 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
450 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
451 struct mlx5e_icosq *sq = &rq->channel->icosq;
452 struct mlx5_wq_cyc *wq = &sq->wq;
453 struct mlx5e_umr_wqe *umr_wqe;
454 u16 xlt_offset = ix << (MLX5E_LOG_ALIGNED_MPWQE_PPW - 1);
455 u16 pi, contig_wqebbs_room;
459 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
460 contig_wqebbs_room = mlx5_wq_cyc_get_contig_wqebbs(wq, pi);
461 if (unlikely(contig_wqebbs_room < MLX5E_UMR_WQEBBS)) {
462 mlx5e_fill_icosq_frag_edge(sq, wq, pi, contig_wqebbs_room);
463 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
466 umr_wqe = mlx5_wq_cyc_get_wqe(wq, pi);
467 if (unlikely(mlx5e_icosq_wrap_cnt(sq) < 2))
468 memcpy(umr_wqe, &rq->mpwqe.umr_wqe,
469 offsetof(struct mlx5e_umr_wqe, inline_mtts));
471 for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++, dma_info++) {
472 err = mlx5e_page_alloc_mapped(rq, dma_info);
475 umr_wqe->inline_mtts[i].ptag = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
478 bitmap_zero(wi->xdp_xmit_bitmap, MLX5_MPWRQ_PAGES_PER_WQE);
479 wi->consumed_strides = 0;
481 rq->mpwqe.umr_in_progress = true;
483 umr_wqe->ctrl.opmod_idx_opcode =
484 cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
486 umr_wqe->uctrl.xlt_offset = cpu_to_be16(xlt_offset);
488 sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_UMR;
489 sq->pc += MLX5E_UMR_WQEBBS;
490 mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &umr_wqe->ctrl);
497 mlx5e_page_release(rq, dma_info, true);
499 rq->stats->buff_alloc_err++;
504 void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
506 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
507 /* Don't recycle, this function is called on rq/netdev close */
508 mlx5e_free_rx_mpwqe(rq, wi, false);
511 bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
513 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
517 if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state)))
520 wqe_bulk = rq->wqe.info.wqe_bulk;
522 if (mlx5_wq_cyc_missing(wq) < wqe_bulk)
526 u16 head = mlx5_wq_cyc_get_head(wq);
528 err = mlx5e_alloc_rx_wqes(rq, head, wqe_bulk);
530 rq->stats->buff_alloc_err++;
534 mlx5_wq_cyc_push_n(wq, wqe_bulk);
535 } while (mlx5_wq_cyc_missing(wq) >= wqe_bulk);
537 /* ensure wqes are visible to device before updating doorbell record */
540 mlx5_wq_cyc_update_db_record(wq);
545 static inline void mlx5e_poll_ico_single_cqe(struct mlx5e_cq *cq,
546 struct mlx5e_icosq *sq,
548 struct mlx5_cqe64 *cqe)
550 struct mlx5_wq_cyc *wq = &sq->wq;
551 u16 ci = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
552 struct mlx5e_sq_wqe_info *icowi = &sq->db.ico_wqe[ci];
554 mlx5_cqwq_pop(&cq->wq);
556 if (unlikely(get_cqe_opcode(cqe) != MLX5_CQE_REQ)) {
557 netdev_WARN_ONCE(cq->channel->netdev,
558 "Bad OP in ICOSQ CQE: 0x%x\n", get_cqe_opcode(cqe));
562 if (likely(icowi->opcode == MLX5_OPCODE_UMR)) {
563 mlx5e_post_rx_mpwqe(rq);
567 if (unlikely(icowi->opcode != MLX5_OPCODE_NOP))
568 netdev_WARN_ONCE(cq->channel->netdev,
569 "Bad OPCODE in ICOSQ WQE info: 0x%x\n", icowi->opcode);
572 static void mlx5e_poll_ico_cq(struct mlx5e_cq *cq, struct mlx5e_rq *rq)
574 struct mlx5e_icosq *sq = container_of(cq, struct mlx5e_icosq, cq);
575 struct mlx5_cqe64 *cqe;
577 if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED, &sq->state)))
580 cqe = mlx5_cqwq_get_cqe(&cq->wq);
584 /* by design, there's only a single cqe */
585 mlx5e_poll_ico_single_cqe(cq, sq, rq, cqe);
587 mlx5_cqwq_update_db_record(&cq->wq);
590 bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq)
592 struct mlx5_wq_ll *wq = &rq->mpwqe.wq;
594 if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state)))
597 mlx5e_poll_ico_cq(&rq->channel->icosq.cq, rq);
599 if (mlx5_wq_ll_is_full(wq))
602 if (!rq->mpwqe.umr_in_progress)
603 mlx5e_alloc_rx_mpwqe(rq, wq->head);
605 rq->stats->congst_umr += mlx5_wq_ll_missing(wq) > 2;
610 static void mlx5e_lro_update_tcp_hdr(struct mlx5_cqe64 *cqe, struct tcphdr *tcp)
612 u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
613 u8 tcp_ack = (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA) ||
614 (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA);
617 tcp->psh = get_cqe_lro_tcppsh(cqe);
621 tcp->ack_seq = cqe->lro_ack_seq_num;
622 tcp->window = cqe->lro_tcp_win;
626 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe,
629 struct ethhdr *eth = (struct ethhdr *)(skb->data);
631 int network_depth = 0;
637 proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
639 tot_len = cqe_bcnt - network_depth;
640 ip_p = skb->data + network_depth;
642 if (proto == htons(ETH_P_IP)) {
643 struct iphdr *ipv4 = ip_p;
645 tcp = ip_p + sizeof(struct iphdr);
646 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
648 ipv4->ttl = cqe->lro_min_ttl;
649 ipv4->tot_len = cpu_to_be16(tot_len);
651 ipv4->check = ip_fast_csum((unsigned char *)ipv4,
654 mlx5e_lro_update_tcp_hdr(cqe, tcp);
655 check = csum_partial(tcp, tcp->doff * 4,
656 csum_unfold((__force __sum16)cqe->check_sum));
657 /* Almost done, don't forget the pseudo header */
658 tcp->check = csum_tcpudp_magic(ipv4->saddr, ipv4->daddr,
659 tot_len - sizeof(struct iphdr),
662 u16 payload_len = tot_len - sizeof(struct ipv6hdr);
663 struct ipv6hdr *ipv6 = ip_p;
665 tcp = ip_p + sizeof(struct ipv6hdr);
666 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
668 ipv6->hop_limit = cqe->lro_min_ttl;
669 ipv6->payload_len = cpu_to_be16(payload_len);
671 mlx5e_lro_update_tcp_hdr(cqe, tcp);
672 check = csum_partial(tcp, tcp->doff * 4,
673 csum_unfold((__force __sum16)cqe->check_sum));
674 /* Almost done, don't forget the pseudo header */
675 tcp->check = csum_ipv6_magic(&ipv6->saddr, &ipv6->daddr, payload_len,
680 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
683 u8 cht = cqe->rss_hash_type;
684 int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
685 (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
687 skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
690 static inline bool is_last_ethertype_ip(struct sk_buff *skb, int *network_depth,
693 *proto = ((struct ethhdr *)skb->data)->h_proto;
694 *proto = __vlan_get_protocol(skb, *proto, network_depth);
695 return (*proto == htons(ETH_P_IP) || *proto == htons(ETH_P_IPV6));
698 static inline void mlx5e_enable_ecn(struct mlx5e_rq *rq, struct sk_buff *skb)
700 int network_depth = 0;
705 if (unlikely(!is_last_ethertype_ip(skb, &network_depth, &proto)))
708 ip = skb->data + network_depth;
709 rc = ((proto == htons(ETH_P_IP)) ? IP_ECN_set_ce((struct iphdr *)ip) :
710 IP6_ECN_set_ce(skb, (struct ipv6hdr *)ip));
712 rq->stats->ecn_mark += !!rc;
715 static u32 mlx5e_get_fcs(const struct sk_buff *skb)
717 const void *fcs_bytes;
720 fcs_bytes = skb_header_pointer(skb, skb->len - ETH_FCS_LEN,
721 ETH_FCS_LEN, &_fcs_bytes);
723 return __get_unaligned_cpu32(fcs_bytes);
726 static u8 get_ip_proto(struct sk_buff *skb, int network_depth, __be16 proto)
728 void *ip_p = skb->data + network_depth;
730 return (proto == htons(ETH_P_IP)) ? ((struct iphdr *)ip_p)->protocol :
731 ((struct ipv6hdr *)ip_p)->nexthdr;
734 #define short_frame(size) ((size) <= ETH_ZLEN + ETH_FCS_LEN)
736 static inline void mlx5e_handle_csum(struct net_device *netdev,
737 struct mlx5_cqe64 *cqe,
742 struct mlx5e_rq_stats *stats = rq->stats;
743 int network_depth = 0;
746 if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
750 skb->ip_summed = CHECKSUM_UNNECESSARY;
751 stats->csum_unnecessary++;
755 if (unlikely(test_bit(MLX5E_RQ_STATE_NO_CSUM_COMPLETE, &rq->state)))
756 goto csum_unnecessary;
758 /* CQE csum doesn't cover padding octets in short ethernet
759 * frames. And the pad field is appended prior to calculating
760 * and appending the FCS field.
762 * Detecting these padded frames requires to verify and parse
763 * IP headers, so we simply force all those small frames to be
764 * CHECKSUM_UNNECESSARY even if they are not padded.
766 if (short_frame(skb->len))
767 goto csum_unnecessary;
769 if (likely(is_last_ethertype_ip(skb, &network_depth, &proto))) {
770 if (unlikely(get_ip_proto(skb, network_depth, proto) == IPPROTO_SCTP))
771 goto csum_unnecessary;
773 skb->ip_summed = CHECKSUM_COMPLETE;
774 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
775 if (network_depth > ETH_HLEN)
776 /* CQE csum is calculated from the IP header and does
777 * not cover VLAN headers (if present). This will add
778 * the checksum manually.
780 skb->csum = csum_partial(skb->data + ETH_HLEN,
781 network_depth - ETH_HLEN,
783 if (unlikely(netdev->features & NETIF_F_RXFCS))
784 skb->csum = csum_block_add(skb->csum,
785 (__force __wsum)mlx5e_get_fcs(skb),
786 skb->len - ETH_FCS_LEN);
787 stats->csum_complete++;
792 if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
793 ((cqe->hds_ip_ext & CQE_L4_OK) ||
794 (get_cqe_l4_hdr_type(cqe) == CQE_L4_HDR_TYPE_NONE)))) {
795 skb->ip_summed = CHECKSUM_UNNECESSARY;
796 if (cqe_is_tunneled(cqe)) {
798 skb->encapsulation = 1;
799 stats->csum_unnecessary_inner++;
802 stats->csum_unnecessary++;
806 skb->ip_summed = CHECKSUM_NONE;
810 #define MLX5E_CE_BIT_MASK 0x80
812 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
817 u8 lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
818 struct mlx5e_rq_stats *stats = rq->stats;
819 struct net_device *netdev = rq->netdev;
821 skb->mac_len = ETH_HLEN;
823 #ifdef CONFIG_MLX5_EN_TLS
824 mlx5e_tls_handle_rx_skb(netdev, skb, &cqe_bcnt);
827 if (lro_num_seg > 1) {
828 mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
829 skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
830 /* Subtract one since we already counted this as one
831 * "regular" packet in mlx5e_complete_rx_cqe()
833 stats->packets += lro_num_seg - 1;
834 stats->lro_packets++;
835 stats->lro_bytes += cqe_bcnt;
838 if (unlikely(mlx5e_rx_hw_stamp(rq->tstamp)))
839 skb_hwtstamps(skb)->hwtstamp =
840 mlx5_timecounter_cyc2time(rq->clock, get_cqe_ts(cqe));
842 skb_record_rx_queue(skb, rq->ix);
844 if (likely(netdev->features & NETIF_F_RXHASH))
845 mlx5e_skb_set_hash(cqe, skb);
847 if (cqe_has_vlan(cqe)) {
848 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
849 be16_to_cpu(cqe->vlan_info));
850 stats->removed_vlan_packets++;
853 skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
855 mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
856 /* checking CE bit in cqe - MSB in ml_path field */
857 if (unlikely(cqe->ml_path & MLX5E_CE_BIT_MASK))
858 mlx5e_enable_ecn(rq, skb);
860 skb->protocol = eth_type_trans(skb, netdev);
863 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
864 struct mlx5_cqe64 *cqe,
868 struct mlx5e_rq_stats *stats = rq->stats;
871 stats->bytes += cqe_bcnt;
872 mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
876 struct sk_buff *mlx5e_build_linear_skb(struct mlx5e_rq *rq, void *va,
877 u32 frag_size, u16 headroom,
880 struct sk_buff *skb = build_skb(va, frag_size);
882 if (unlikely(!skb)) {
883 rq->stats->buff_alloc_err++;
887 skb_reserve(skb, headroom);
888 skb_put(skb, cqe_bcnt);
894 mlx5e_skb_from_cqe_linear(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
895 struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
897 struct mlx5e_dma_info *di = wi->di;
898 u16 rx_headroom = rq->buff.headroom;
904 va = page_address(di->page) + wi->offset;
905 data = va + rx_headroom;
906 frag_size = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
908 dma_sync_single_range_for_cpu(rq->pdev, di->addr, wi->offset,
909 frag_size, DMA_FROM_DEVICE);
910 prefetchw(va); /* xdp_frame data area */
913 if (unlikely(get_cqe_opcode(cqe) != MLX5_CQE_RESP_SEND)) {
914 rq->stats->wqe_err++;
919 consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt);
922 return NULL; /* page/packet was consumed by XDP */
924 skb = mlx5e_build_linear_skb(rq, va, frag_size, rx_headroom, cqe_bcnt);
928 /* queue up for recycling/reuse */
929 page_ref_inc(di->page);
935 mlx5e_skb_from_cqe_nonlinear(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
936 struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
938 struct mlx5e_rq_frag_info *frag_info = &rq->wqe.info.arr[0];
939 struct mlx5e_wqe_frag_info *head_wi = wi;
940 u16 headlen = min_t(u32, MLX5E_RX_MAX_HEAD, cqe_bcnt);
941 u16 frag_headlen = headlen;
942 u16 byte_cnt = cqe_bcnt - headlen;
945 if (unlikely(get_cqe_opcode(cqe) != MLX5_CQE_RESP_SEND)) {
946 rq->stats->wqe_err++;
950 /* XDP is not supported in this configuration, as incoming packets
951 * might spread among multiple pages.
953 skb = napi_alloc_skb(rq->cq.napi,
954 ALIGN(MLX5E_RX_MAX_HEAD, sizeof(long)));
955 if (unlikely(!skb)) {
956 rq->stats->buff_alloc_err++;
960 prefetchw(skb->data);
963 u16 frag_consumed_bytes =
964 min_t(u16, frag_info->frag_size - frag_headlen, byte_cnt);
966 mlx5e_add_skb_frag(rq, skb, wi->di, wi->offset + frag_headlen,
967 frag_consumed_bytes, frag_info->frag_stride);
968 byte_cnt -= frag_consumed_bytes;
975 mlx5e_copy_skb_header(rq->pdev, skb, head_wi->di, head_wi->offset, headlen);
976 /* skb linear part was allocated with headlen and aligned to long */
977 skb->tail += headlen;
983 void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
985 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
986 struct mlx5e_wqe_frag_info *wi;
991 ci = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
992 wi = get_frag(rq, ci);
993 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
995 skb = rq->wqe.skb_from_cqe(rq, cqe, wi, cqe_bcnt);
997 /* probably for XDP */
998 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)) {
999 /* do not return page to cache,
1000 * it will be returned on XDP_TX completion.
1007 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1008 napi_gro_receive(rq->cq.napi, skb);
1011 mlx5e_free_rx_wqe(rq, wi, true);
1013 mlx5_wq_cyc_pop(wq);
1016 #ifdef CONFIG_MLX5_ESWITCH
1017 void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1019 struct net_device *netdev = rq->netdev;
1020 struct mlx5e_priv *priv = netdev_priv(netdev);
1021 struct mlx5e_rep_priv *rpriv = priv->ppriv;
1022 struct mlx5_eswitch_rep *rep = rpriv->rep;
1023 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1024 struct mlx5e_wqe_frag_info *wi;
1025 struct sk_buff *skb;
1029 ci = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1030 wi = get_frag(rq, ci);
1031 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1033 skb = rq->wqe.skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1035 /* probably for XDP */
1036 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)) {
1037 /* do not return page to cache,
1038 * it will be returned on XDP_TX completion.
1045 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1047 if (rep->vlan && skb_vlan_tag_present(skb))
1050 napi_gro_receive(rq->cq.napi, skb);
1053 mlx5e_free_rx_wqe(rq, wi, true);
1055 mlx5_wq_cyc_pop(wq);
1060 mlx5e_skb_from_cqe_mpwrq_nonlinear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
1061 u16 cqe_bcnt, u32 head_offset, u32 page_idx)
1063 u16 headlen = min_t(u16, MLX5E_RX_MAX_HEAD, cqe_bcnt);
1064 struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
1065 u32 frag_offset = head_offset + headlen;
1066 u32 byte_cnt = cqe_bcnt - headlen;
1067 struct mlx5e_dma_info *head_di = di;
1068 struct sk_buff *skb;
1070 skb = napi_alloc_skb(rq->cq.napi,
1071 ALIGN(MLX5E_RX_MAX_HEAD, sizeof(long)));
1072 if (unlikely(!skb)) {
1073 rq->stats->buff_alloc_err++;
1077 prefetchw(skb->data);
1079 if (unlikely(frag_offset >= PAGE_SIZE)) {
1081 frag_offset -= PAGE_SIZE;
1085 u32 pg_consumed_bytes =
1086 min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
1087 unsigned int truesize =
1088 ALIGN(pg_consumed_bytes, BIT(rq->mpwqe.log_stride_sz));
1090 mlx5e_add_skb_frag(rq, skb, di, frag_offset,
1091 pg_consumed_bytes, truesize);
1092 byte_cnt -= pg_consumed_bytes;
1097 mlx5e_copy_skb_header(rq->pdev, skb, head_di, head_offset, headlen);
1098 /* skb linear part was allocated with headlen and aligned to long */
1099 skb->tail += headlen;
1100 skb->len += headlen;
1106 mlx5e_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi,
1107 u16 cqe_bcnt, u32 head_offset, u32 page_idx)
1109 struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
1110 u16 rx_headroom = rq->buff.headroom;
1111 u32 cqe_bcnt32 = cqe_bcnt;
1112 struct sk_buff *skb;
1117 /* Check packet size. Note LRO doesn't use linear SKB */
1118 if (unlikely(cqe_bcnt > rq->hw_mtu)) {
1119 rq->stats->oversize_pkts_sw_drop++;
1123 va = page_address(di->page) + head_offset;
1124 data = va + rx_headroom;
1125 frag_size = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt32);
1127 dma_sync_single_range_for_cpu(rq->pdev, di->addr, head_offset,
1128 frag_size, DMA_FROM_DEVICE);
1129 prefetchw(va); /* xdp_frame data area */
1133 consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt32);
1136 if (__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags))
1137 __set_bit(page_idx, wi->xdp_xmit_bitmap); /* non-atomic */
1138 return NULL; /* page/packet was consumed by XDP */
1141 skb = mlx5e_build_linear_skb(rq, va, frag_size, rx_headroom, cqe_bcnt32);
1145 /* queue up for recycling/reuse */
1146 page_ref_inc(di->page);
1151 void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1153 u16 cstrides = mpwrq_get_cqe_consumed_strides(cqe);
1154 u16 wqe_id = be16_to_cpu(cqe->wqe_id);
1155 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
1156 u16 stride_ix = mpwrq_get_cqe_stride_index(cqe);
1157 u32 wqe_offset = stride_ix << rq->mpwqe.log_stride_sz;
1158 u32 head_offset = wqe_offset & (PAGE_SIZE - 1);
1159 u32 page_idx = wqe_offset >> PAGE_SHIFT;
1160 struct mlx5e_rx_wqe_ll *wqe;
1161 struct mlx5_wq_ll *wq;
1162 struct sk_buff *skb;
1165 wi->consumed_strides += cstrides;
1167 if (unlikely(get_cqe_opcode(cqe) != MLX5_CQE_RESP_SEND)) {
1168 rq->stats->wqe_err++;
1172 if (unlikely(mpwrq_is_filler_cqe(cqe))) {
1173 struct mlx5e_rq_stats *stats = rq->stats;
1175 stats->mpwqe_filler_cqes++;
1176 stats->mpwqe_filler_strides += cstrides;
1180 cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
1182 skb = rq->mpwqe.skb_from_cqe_mpwrq(rq, wi, cqe_bcnt, head_offset,
1187 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1188 napi_gro_receive(rq->cq.napi, skb);
1191 if (likely(wi->consumed_strides < rq->mpwqe.num_strides))
1195 wqe = mlx5_wq_ll_get_wqe(wq, wqe_id);
1196 mlx5e_free_rx_mpwqe(rq, wi, true);
1197 mlx5_wq_ll_pop(wq, cqe->wqe_id, &wqe->next.next_wqe_index);
1200 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
1202 struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
1203 struct mlx5_cqwq *cqwq = &cq->wq;
1204 struct mlx5_cqe64 *cqe;
1207 if (unlikely(!test_bit(MLX5E_RQ_STATE_ENABLED, &rq->state)))
1211 work_done += mlx5e_decompress_cqes_cont(rq, cqwq, 0, budget);
1213 cqe = mlx5_cqwq_get_cqe(cqwq);
1215 if (unlikely(work_done))
1221 if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
1223 mlx5e_decompress_cqes_start(rq, cqwq,
1224 budget - work_done);
1228 mlx5_cqwq_pop(cqwq);
1230 rq->handle_rx_cqe(rq, cqe);
1231 } while ((++work_done < budget) && (cqe = mlx5_cqwq_get_cqe(cqwq)));
1235 mlx5e_xdp_rx_poll_complete(rq);
1237 mlx5_cqwq_update_db_record(cqwq);
1239 /* ensure cq space is freed before enabling more cqes */
1245 #ifdef CONFIG_MLX5_CORE_IPOIB
1247 #define MLX5_IB_GRH_DGID_OFFSET 24
1248 #define MLX5_GID_SIZE 16
1250 static inline void mlx5i_complete_rx_cqe(struct mlx5e_rq *rq,
1251 struct mlx5_cqe64 *cqe,
1253 struct sk_buff *skb)
1255 struct hwtstamp_config *tstamp;
1256 struct mlx5e_rq_stats *stats;
1257 struct net_device *netdev;
1258 struct mlx5e_priv *priv;
1259 char *pseudo_header;
1264 qpn = be32_to_cpu(cqe->sop_drop_qpn) & 0xffffff;
1265 netdev = mlx5i_pkey_get_netdev(rq->netdev, qpn);
1267 /* No mapping present, cannot process SKB. This might happen if a child
1268 * interface is going down while having unprocessed CQEs on parent RQ
1270 if (unlikely(!netdev)) {
1271 /* TODO: add drop counters support */
1273 pr_warn_once("Unable to map QPN %u to dev - dropping skb\n", qpn);
1277 priv = mlx5i_epriv(netdev);
1278 tstamp = &priv->tstamp;
1279 stats = &priv->channel_stats[rq->ix].rq;
1281 g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
1282 dgid = skb->data + MLX5_IB_GRH_DGID_OFFSET;
1283 if ((!g) || dgid[0] != 0xff)
1284 skb->pkt_type = PACKET_HOST;
1285 else if (memcmp(dgid, netdev->broadcast + 4, MLX5_GID_SIZE) == 0)
1286 skb->pkt_type = PACKET_BROADCAST;
1288 skb->pkt_type = PACKET_MULTICAST;
1290 /* TODO: IB/ipoib: Allow mcast packets from other VFs
1291 * 68996a6e760e5c74654723eeb57bf65628ae87f4
1294 skb_pull(skb, MLX5_IB_GRH_BYTES);
1296 skb->protocol = *((__be16 *)(skb->data));
1298 if (netdev->features & NETIF_F_RXCSUM) {
1299 skb->ip_summed = CHECKSUM_COMPLETE;
1300 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1301 stats->csum_complete++;
1303 skb->ip_summed = CHECKSUM_NONE;
1307 if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
1308 skb_hwtstamps(skb)->hwtstamp =
1309 mlx5_timecounter_cyc2time(rq->clock, get_cqe_ts(cqe));
1311 skb_record_rx_queue(skb, rq->ix);
1313 if (likely(netdev->features & NETIF_F_RXHASH))
1314 mlx5e_skb_set_hash(cqe, skb);
1316 /* 20 bytes of ipoib header and 4 for encap existing */
1317 pseudo_header = skb_push(skb, MLX5_IPOIB_PSEUDO_LEN);
1318 memset(pseudo_header, 0, MLX5_IPOIB_PSEUDO_LEN);
1319 skb_reset_mac_header(skb);
1320 skb_pull(skb, MLX5_IPOIB_HARD_LEN);
1325 stats->bytes += cqe_bcnt;
1328 void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1330 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1331 struct mlx5e_wqe_frag_info *wi;
1332 struct sk_buff *skb;
1336 ci = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1337 wi = get_frag(rq, ci);
1338 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1340 skb = rq->wqe.skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1344 mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1345 if (unlikely(!skb->dev)) {
1346 dev_kfree_skb_any(skb);
1349 napi_gro_receive(rq->cq.napi, skb);
1352 mlx5e_free_rx_wqe(rq, wi, true);
1353 mlx5_wq_cyc_pop(wq);
1356 #endif /* CONFIG_MLX5_CORE_IPOIB */
1358 #ifdef CONFIG_MLX5_EN_IPSEC
1360 void mlx5e_ipsec_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1362 struct mlx5_wq_cyc *wq = &rq->wqe.wq;
1363 struct mlx5e_wqe_frag_info *wi;
1364 struct sk_buff *skb;
1368 ci = mlx5_wq_cyc_ctr2ix(wq, be16_to_cpu(cqe->wqe_counter));
1369 wi = get_frag(rq, ci);
1370 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1372 skb = rq->wqe.skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1373 if (unlikely(!skb)) {
1374 /* a DROP, save the page-reuse checks */
1375 mlx5e_free_rx_wqe(rq, wi, true);
1378 skb = mlx5e_ipsec_handle_rx_skb(rq->netdev, skb, &cqe_bcnt);
1379 if (unlikely(!skb)) {
1380 mlx5e_free_rx_wqe(rq, wi, true);
1384 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1385 napi_gro_receive(rq->cq.napi, skb);
1387 mlx5e_free_rx_wqe(rq, wi, true);
1389 mlx5_wq_cyc_pop(wq);
1392 #endif /* CONFIG_MLX5_EN_IPSEC */