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 <linux/bpf_trace.h>
38 #include <net/busy_poll.h>
43 #include "ipoib/ipoib.h"
44 #include "en_accel/ipsec_rxtx.h"
45 #include "lib/clock.h"
47 static inline bool mlx5e_rx_hw_stamp(struct hwtstamp_config *config)
49 return config->rx_filter == HWTSTAMP_FILTER_ALL;
52 static inline void mlx5e_read_cqe_slot(struct mlx5e_cq *cq, u32 cqcc,
55 u32 ci = cqcc & cq->wq.sz_m1;
57 memcpy(data, mlx5_cqwq_get_wqe(&cq->wq, ci), sizeof(struct mlx5_cqe64));
60 static inline void mlx5e_read_title_slot(struct mlx5e_rq *rq,
61 struct mlx5e_cq *cq, u32 cqcc)
63 mlx5e_read_cqe_slot(cq, cqcc, &cq->title);
64 cq->decmprs_left = be32_to_cpu(cq->title.byte_cnt);
65 cq->decmprs_wqe_counter = be16_to_cpu(cq->title.wqe_counter);
66 rq->stats.cqe_compress_blks++;
69 static inline void mlx5e_read_mini_arr_slot(struct mlx5e_cq *cq, u32 cqcc)
71 mlx5e_read_cqe_slot(cq, cqcc, cq->mini_arr);
75 static inline void mlx5e_cqes_update_owner(struct mlx5e_cq *cq, u32 cqcc, int n)
77 u8 op_own = (cqcc >> cq->wq.log_sz) & 1;
78 u32 wq_sz = 1 << cq->wq.log_sz;
79 u32 ci = cqcc & cq->wq.sz_m1;
80 u32 ci_top = min_t(u32, wq_sz, ci + n);
82 for (; ci < ci_top; ci++, n--) {
83 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
88 if (unlikely(ci == wq_sz)) {
90 for (ci = 0; ci < n; ci++) {
91 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, ci);
98 static inline void mlx5e_decompress_cqe(struct mlx5e_rq *rq,
99 struct mlx5e_cq *cq, u32 cqcc)
101 cq->title.byte_cnt = cq->mini_arr[cq->mini_arr_idx].byte_cnt;
102 cq->title.check_sum = cq->mini_arr[cq->mini_arr_idx].checksum;
103 cq->title.op_own &= 0xf0;
104 cq->title.op_own |= 0x01 & (cqcc >> cq->wq.log_sz);
105 cq->title.wqe_counter = cpu_to_be16(cq->decmprs_wqe_counter);
107 if (rq->wq_type == MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ)
108 cq->decmprs_wqe_counter +=
109 mpwrq_get_cqe_consumed_strides(&cq->title);
111 cq->decmprs_wqe_counter =
112 (cq->decmprs_wqe_counter + 1) & rq->wq.sz_m1;
115 static inline void mlx5e_decompress_cqe_no_hash(struct mlx5e_rq *rq,
116 struct mlx5e_cq *cq, u32 cqcc)
118 mlx5e_decompress_cqe(rq, cq, cqcc);
119 cq->title.rss_hash_type = 0;
120 cq->title.rss_hash_result = 0;
123 static inline u32 mlx5e_decompress_cqes_cont(struct mlx5e_rq *rq,
125 int update_owner_only,
128 u32 cqcc = cq->wq.cc + update_owner_only;
132 cqe_count = min_t(u32, cq->decmprs_left, budget_rem);
134 for (i = update_owner_only; i < cqe_count;
135 i++, cq->mini_arr_idx++, cqcc++) {
136 if (cq->mini_arr_idx == MLX5_MINI_CQE_ARRAY_SIZE)
137 mlx5e_read_mini_arr_slot(cq, cqcc);
139 mlx5e_decompress_cqe_no_hash(rq, cq, cqcc);
140 rq->handle_rx_cqe(rq, &cq->title);
142 mlx5e_cqes_update_owner(cq, cq->wq.cc, cqcc - cq->wq.cc);
144 cq->decmprs_left -= cqe_count;
145 rq->stats.cqe_compress_pkts += cqe_count;
150 static inline u32 mlx5e_decompress_cqes_start(struct mlx5e_rq *rq,
154 mlx5e_read_title_slot(rq, cq, cq->wq.cc);
155 mlx5e_read_mini_arr_slot(cq, cq->wq.cc + 1);
156 mlx5e_decompress_cqe(rq, cq, cq->wq.cc);
157 rq->handle_rx_cqe(rq, &cq->title);
160 return mlx5e_decompress_cqes_cont(rq, cq, 1, budget_rem) - 1;
163 #define RQ_PAGE_SIZE(rq) ((1 << rq->buff.page_order) << PAGE_SHIFT)
165 static inline bool mlx5e_page_is_reserved(struct page *page)
167 return page_is_pfmemalloc(page) || page_to_nid(page) != numa_mem_id();
170 static inline bool mlx5e_rx_cache_put(struct mlx5e_rq *rq,
171 struct mlx5e_dma_info *dma_info)
173 struct mlx5e_page_cache *cache = &rq->page_cache;
174 u32 tail_next = (cache->tail + 1) & (MLX5E_CACHE_SIZE - 1);
176 if (tail_next == cache->head) {
177 rq->stats.cache_full++;
181 if (unlikely(mlx5e_page_is_reserved(dma_info->page))) {
182 rq->stats.cache_waive++;
186 cache->page_cache[cache->tail] = *dma_info;
187 cache->tail = tail_next;
191 static inline bool mlx5e_rx_cache_get(struct mlx5e_rq *rq,
192 struct mlx5e_dma_info *dma_info)
194 struct mlx5e_page_cache *cache = &rq->page_cache;
196 if (unlikely(cache->head == cache->tail)) {
197 rq->stats.cache_empty++;
201 if (page_ref_count(cache->page_cache[cache->head].page) != 1) {
202 rq->stats.cache_busy++;
206 *dma_info = cache->page_cache[cache->head];
207 cache->head = (cache->head + 1) & (MLX5E_CACHE_SIZE - 1);
208 rq->stats.cache_reuse++;
210 dma_sync_single_for_device(rq->pdev, dma_info->addr,
216 static inline int mlx5e_page_alloc_mapped(struct mlx5e_rq *rq,
217 struct mlx5e_dma_info *dma_info)
219 if (mlx5e_rx_cache_get(rq, dma_info))
222 dma_info->page = dev_alloc_pages(rq->buff.page_order);
223 if (unlikely(!dma_info->page))
226 dma_info->addr = dma_map_page(rq->pdev, dma_info->page, 0,
227 RQ_PAGE_SIZE(rq), rq->buff.map_dir);
228 if (unlikely(dma_mapping_error(rq->pdev, dma_info->addr))) {
229 put_page(dma_info->page);
230 dma_info->page = NULL;
237 void mlx5e_page_release(struct mlx5e_rq *rq, struct mlx5e_dma_info *dma_info,
240 if (likely(recycle) && mlx5e_rx_cache_put(rq, dma_info))
243 dma_unmap_page(rq->pdev, dma_info->addr, RQ_PAGE_SIZE(rq),
245 put_page(dma_info->page);
248 static inline bool mlx5e_page_reuse(struct mlx5e_rq *rq,
249 struct mlx5e_wqe_frag_info *wi)
251 return rq->wqe.page_reuse && wi->di.page &&
252 (wi->offset + rq->wqe.frag_sz <= RQ_PAGE_SIZE(rq)) &&
253 !mlx5e_page_is_reserved(wi->di.page);
256 static int mlx5e_alloc_rx_wqe(struct mlx5e_rq *rq, struct mlx5e_rx_wqe *wqe, u16 ix)
258 struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
260 /* check if page exists, hence can be reused */
262 if (unlikely(mlx5e_page_alloc_mapped(rq, &wi->di)))
267 wqe->data.addr = cpu_to_be64(wi->di.addr + wi->offset + rq->buff.headroom);
271 static inline void mlx5e_free_rx_wqe(struct mlx5e_rq *rq,
272 struct mlx5e_wqe_frag_info *wi)
274 mlx5e_page_release(rq, &wi->di, true);
278 static inline void mlx5e_free_rx_wqe_reuse(struct mlx5e_rq *rq,
279 struct mlx5e_wqe_frag_info *wi)
281 if (mlx5e_page_reuse(rq, wi)) {
282 rq->stats.page_reuse++;
286 mlx5e_free_rx_wqe(rq, wi);
289 void mlx5e_dealloc_rx_wqe(struct mlx5e_rq *rq, u16 ix)
291 struct mlx5e_wqe_frag_info *wi = &rq->wqe.frag_info[ix];
294 mlx5e_free_rx_wqe(rq, wi);
297 static inline int mlx5e_mpwqe_strides_per_page(struct mlx5e_rq *rq)
299 return rq->mpwqe.num_strides >> MLX5_MPWRQ_WQE_PAGE_ORDER;
302 static inline void mlx5e_add_skb_frag_mpwqe(struct mlx5e_rq *rq,
304 struct mlx5e_mpw_info *wi,
305 u32 page_idx, u32 frag_offset,
308 unsigned int truesize = ALIGN(len, BIT(rq->mpwqe.log_stride_sz));
310 dma_sync_single_for_cpu(rq->pdev,
311 wi->umr.dma_info[page_idx].addr + frag_offset,
312 len, DMA_FROM_DEVICE);
313 wi->skbs_frags[page_idx]++;
314 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
315 wi->umr.dma_info[page_idx].page, frag_offset,
320 mlx5e_copy_skb_header_mpwqe(struct device *pdev,
322 struct mlx5e_mpw_info *wi,
323 u32 page_idx, u32 offset,
326 u16 headlen_pg = min_t(u32, headlen, PAGE_SIZE - offset);
327 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[page_idx];
330 /* Aligning len to sizeof(long) optimizes memcpy performance */
331 len = ALIGN(headlen_pg, sizeof(long));
332 dma_sync_single_for_cpu(pdev, dma_info->addr + offset, len,
334 skb_copy_to_linear_data_offset(skb, 0,
335 page_address(dma_info->page) + offset,
337 if (unlikely(offset + headlen > PAGE_SIZE)) {
340 len = ALIGN(headlen - headlen_pg, sizeof(long));
341 dma_sync_single_for_cpu(pdev, dma_info->addr, len,
343 skb_copy_to_linear_data_offset(skb, headlen_pg,
344 page_address(dma_info->page),
349 static inline void mlx5e_post_umr_wqe(struct mlx5e_rq *rq, u16 ix)
351 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
352 struct mlx5e_icosq *sq = &rq->channel->icosq;
353 struct mlx5_wq_cyc *wq = &sq->wq;
354 struct mlx5e_umr_wqe *wqe;
355 u8 num_wqebbs = DIV_ROUND_UP(sizeof(*wqe), MLX5_SEND_WQE_BB);
358 /* fill sq edge with nops to avoid wqe wrap around */
359 while ((pi = (sq->pc & wq->sz_m1)) > sq->edge) {
360 sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_NOP;
361 mlx5e_post_nop(wq, sq->sqn, &sq->pc);
364 wqe = mlx5_wq_cyc_get_wqe(wq, pi);
365 memcpy(wqe, &wi->umr.wqe, sizeof(*wqe));
366 wqe->ctrl.opmod_idx_opcode =
367 cpu_to_be32((sq->pc << MLX5_WQE_CTRL_WQE_INDEX_SHIFT) |
370 sq->db.ico_wqe[pi].opcode = MLX5_OPCODE_UMR;
371 sq->pc += num_wqebbs;
372 mlx5e_notify_hw(&sq->wq, sq->pc, sq->uar_map, &wqe->ctrl);
375 static int mlx5e_alloc_rx_umr_mpwqe(struct mlx5e_rq *rq,
378 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
379 int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
380 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
384 for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++, dma_info++) {
385 err = mlx5e_page_alloc_mapped(rq, dma_info);
388 wi->umr.mtt[i] = cpu_to_be64(dma_info->addr | MLX5_EN_WR);
389 page_ref_add(dma_info->page, pg_strides);
392 memset(wi->skbs_frags, 0, sizeof(*wi->skbs_frags) * MLX5_MPWRQ_PAGES_PER_WQE);
393 wi->consumed_strides = 0;
400 page_ref_sub(dma_info->page, pg_strides);
401 mlx5e_page_release(rq, dma_info, true);
407 void mlx5e_free_rx_mpwqe(struct mlx5e_rq *rq, struct mlx5e_mpw_info *wi)
409 int pg_strides = mlx5e_mpwqe_strides_per_page(rq);
410 struct mlx5e_dma_info *dma_info = &wi->umr.dma_info[0];
413 for (i = 0; i < MLX5_MPWRQ_PAGES_PER_WQE; i++, dma_info++) {
414 page_ref_sub(dma_info->page, pg_strides - wi->skbs_frags[i]);
415 mlx5e_page_release(rq, dma_info, true);
419 static void mlx5e_post_rx_mpwqe(struct mlx5e_rq *rq)
421 struct mlx5_wq_ll *wq = &rq->wq;
422 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
424 rq->mpwqe.umr_in_progress = false;
426 mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
428 /* ensure wqes are visible to device before updating doorbell record */
431 mlx5_wq_ll_update_db_record(wq);
434 static int mlx5e_alloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
438 err = mlx5e_alloc_rx_umr_mpwqe(rq, ix);
440 rq->stats.buff_alloc_err++;
443 rq->mpwqe.umr_in_progress = true;
444 mlx5e_post_umr_wqe(rq, ix);
448 void mlx5e_dealloc_rx_mpwqe(struct mlx5e_rq *rq, u16 ix)
450 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[ix];
452 mlx5e_free_rx_mpwqe(rq, wi);
455 bool mlx5e_post_rx_wqes(struct mlx5e_rq *rq)
457 struct mlx5_wq_ll *wq = &rq->wq;
460 if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
463 if (mlx5_wq_ll_is_full(wq))
467 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(wq, wq->head);
469 err = mlx5e_alloc_rx_wqe(rq, wqe, wq->head);
471 rq->stats.buff_alloc_err++;
475 mlx5_wq_ll_push(wq, be16_to_cpu(wqe->next.next_wqe_index));
476 } while (!mlx5_wq_ll_is_full(wq));
478 /* ensure wqes are visible to device before updating doorbell record */
481 mlx5_wq_ll_update_db_record(wq);
486 static inline void mlx5e_poll_ico_single_cqe(struct mlx5e_cq *cq,
487 struct mlx5e_icosq *sq,
489 struct mlx5_cqe64 *cqe)
491 struct mlx5_wq_cyc *wq = &sq->wq;
492 u16 ci = be16_to_cpu(cqe->wqe_counter) & wq->sz_m1;
493 struct mlx5e_sq_wqe_info *icowi = &sq->db.ico_wqe[ci];
495 mlx5_cqwq_pop(&cq->wq);
497 if (unlikely((cqe->op_own >> 4) != MLX5_CQE_REQ)) {
498 netdev_WARN_ONCE(cq->channel->netdev,
499 "Bad OP in ICOSQ CQE: 0x%x\n", cqe->op_own);
503 if (likely(icowi->opcode == MLX5_OPCODE_UMR)) {
504 mlx5e_post_rx_mpwqe(rq);
508 if (unlikely(icowi->opcode != MLX5_OPCODE_NOP))
509 netdev_WARN_ONCE(cq->channel->netdev,
510 "Bad OPCODE in ICOSQ WQE info: 0x%x\n", icowi->opcode);
513 static void mlx5e_poll_ico_cq(struct mlx5e_cq *cq, struct mlx5e_rq *rq)
515 struct mlx5e_icosq *sq = container_of(cq, struct mlx5e_icosq, cq);
516 struct mlx5_cqe64 *cqe;
518 if (unlikely(!MLX5E_TEST_BIT(sq->state, MLX5E_SQ_STATE_ENABLED)))
521 cqe = mlx5_cqwq_get_cqe(&cq->wq);
525 /* by design, there's only a single cqe */
526 mlx5e_poll_ico_single_cqe(cq, sq, rq, cqe);
528 mlx5_cqwq_update_db_record(&cq->wq);
531 bool mlx5e_post_rx_mpwqes(struct mlx5e_rq *rq)
533 struct mlx5_wq_ll *wq = &rq->wq;
535 if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
538 mlx5e_poll_ico_cq(&rq->channel->icosq.cq, rq);
540 if (mlx5_wq_ll_is_full(wq))
543 if (!rq->mpwqe.umr_in_progress)
544 mlx5e_alloc_rx_mpwqe(rq, wq->head);
549 static void mlx5e_lro_update_hdr(struct sk_buff *skb, struct mlx5_cqe64 *cqe,
552 struct ethhdr *eth = (struct ethhdr *)(skb->data);
554 int network_depth = 0;
559 u8 l4_hdr_type = get_cqe_l4_hdr_type(cqe);
560 u8 tcp_ack = (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_NO_DATA) ||
561 (l4_hdr_type == CQE_L4_HDR_TYPE_TCP_ACK_AND_DATA);
563 proto = __vlan_get_protocol(skb, eth->h_proto, &network_depth);
565 tot_len = cqe_bcnt - network_depth;
566 ip_p = skb->data + network_depth;
568 if (proto == htons(ETH_P_IP)) {
569 struct iphdr *ipv4 = ip_p;
571 tcp = ip_p + sizeof(struct iphdr);
572 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
574 ipv4->ttl = cqe->lro_min_ttl;
575 ipv4->tot_len = cpu_to_be16(tot_len);
577 ipv4->check = ip_fast_csum((unsigned char *)ipv4,
580 struct ipv6hdr *ipv6 = ip_p;
582 tcp = ip_p + sizeof(struct ipv6hdr);
583 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
585 ipv6->hop_limit = cqe->lro_min_ttl;
586 ipv6->payload_len = cpu_to_be16(tot_len -
587 sizeof(struct ipv6hdr));
590 tcp->psh = get_cqe_lro_tcppsh(cqe);
594 tcp->ack_seq = cqe->lro_ack_seq_num;
595 tcp->window = cqe->lro_tcp_win;
599 static inline void mlx5e_skb_set_hash(struct mlx5_cqe64 *cqe,
602 u8 cht = cqe->rss_hash_type;
603 int ht = (cht & CQE_RSS_HTYPE_L4) ? PKT_HASH_TYPE_L4 :
604 (cht & CQE_RSS_HTYPE_IP) ? PKT_HASH_TYPE_L3 :
606 skb_set_hash(skb, be32_to_cpu(cqe->rss_hash_result), ht);
609 static inline bool is_last_ethertype_ip(struct sk_buff *skb, int *network_depth)
611 __be16 ethertype = ((struct ethhdr *)skb->data)->h_proto;
613 ethertype = __vlan_get_protocol(skb, ethertype, network_depth);
614 return (ethertype == htons(ETH_P_IP) || ethertype == htons(ETH_P_IPV6));
617 static inline void mlx5e_handle_csum(struct net_device *netdev,
618 struct mlx5_cqe64 *cqe,
623 int network_depth = 0;
625 if (unlikely(!(netdev->features & NETIF_F_RXCSUM)))
629 skb->ip_summed = CHECKSUM_UNNECESSARY;
630 rq->stats.csum_unnecessary++;
634 if (likely(is_last_ethertype_ip(skb, &network_depth))) {
635 skb->ip_summed = CHECKSUM_COMPLETE;
636 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
637 if (network_depth > ETH_HLEN)
638 /* CQE csum is calculated from the IP header and does
639 * not cover VLAN headers (if present). This will add
640 * the checksum manually.
642 skb->csum = csum_partial(skb->data + ETH_HLEN,
643 network_depth - ETH_HLEN,
645 rq->stats.csum_complete++;
649 if (likely((cqe->hds_ip_ext & CQE_L3_OK) &&
650 (cqe->hds_ip_ext & CQE_L4_OK))) {
651 skb->ip_summed = CHECKSUM_UNNECESSARY;
652 if (cqe_is_tunneled(cqe)) {
654 skb->encapsulation = 1;
655 rq->stats.csum_unnecessary_inner++;
658 rq->stats.csum_unnecessary++;
662 skb->ip_summed = CHECKSUM_NONE;
663 rq->stats.csum_none++;
666 static inline void mlx5e_build_rx_skb(struct mlx5_cqe64 *cqe,
671 struct net_device *netdev = rq->netdev;
674 skb->mac_len = ETH_HLEN;
675 lro_num_seg = be32_to_cpu(cqe->srqn) >> 24;
676 if (lro_num_seg > 1) {
677 mlx5e_lro_update_hdr(skb, cqe, cqe_bcnt);
678 skb_shinfo(skb)->gso_size = DIV_ROUND_UP(cqe_bcnt, lro_num_seg);
679 /* Subtract one since we already counted this as one
680 * "regular" packet in mlx5e_complete_rx_cqe()
682 rq->stats.packets += lro_num_seg - 1;
683 rq->stats.lro_packets++;
684 rq->stats.lro_bytes += cqe_bcnt;
687 if (unlikely(mlx5e_rx_hw_stamp(rq->tstamp)))
688 skb_hwtstamps(skb)->hwtstamp =
689 mlx5_timecounter_cyc2time(rq->clock, get_cqe_ts(cqe));
691 skb_record_rx_queue(skb, rq->ix);
693 if (likely(netdev->features & NETIF_F_RXHASH))
694 mlx5e_skb_set_hash(cqe, skb);
696 if (cqe_has_vlan(cqe)) {
697 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
698 be16_to_cpu(cqe->vlan_info));
699 rq->stats.removed_vlan_packets++;
702 skb->mark = be32_to_cpu(cqe->sop_drop_qpn) & MLX5E_TC_FLOW_ID_MASK;
704 mlx5e_handle_csum(netdev, cqe, rq, skb, !!lro_num_seg);
705 skb->protocol = eth_type_trans(skb, netdev);
708 static inline void mlx5e_complete_rx_cqe(struct mlx5e_rq *rq,
709 struct mlx5_cqe64 *cqe,
714 rq->stats.bytes += cqe_bcnt;
715 mlx5e_build_rx_skb(cqe, cqe_bcnt, rq, skb);
718 static inline void mlx5e_xmit_xdp_doorbell(struct mlx5e_xdpsq *sq)
720 struct mlx5_wq_cyc *wq = &sq->wq;
721 struct mlx5e_tx_wqe *wqe;
722 u16 pi = (sq->pc - 1) & wq->sz_m1; /* last pi */
724 wqe = mlx5_wq_cyc_get_wqe(wq, pi);
726 mlx5e_notify_hw(wq, sq->pc, sq->uar_map, &wqe->ctrl);
729 static inline bool mlx5e_xmit_xdp_frame(struct mlx5e_rq *rq,
730 struct mlx5e_dma_info *di,
731 const struct xdp_buff *xdp)
733 struct mlx5e_xdpsq *sq = &rq->xdpsq;
734 struct mlx5_wq_cyc *wq = &sq->wq;
735 u16 pi = sq->pc & wq->sz_m1;
736 struct mlx5e_tx_wqe *wqe = mlx5_wq_cyc_get_wqe(wq, pi);
738 struct mlx5_wqe_ctrl_seg *cseg = &wqe->ctrl;
739 struct mlx5_wqe_eth_seg *eseg = &wqe->eth;
740 struct mlx5_wqe_data_seg *dseg;
742 ptrdiff_t data_offset = xdp->data - xdp->data_hard_start;
743 dma_addr_t dma_addr = di->addr + data_offset;
744 unsigned int dma_len = xdp->data_end - xdp->data;
748 if (unlikely(dma_len < MLX5E_XDP_MIN_INLINE ||
749 MLX5E_SW2HW_MTU(rq->channel->priv, rq->netdev->mtu) < dma_len)) {
750 rq->stats.xdp_drop++;
754 if (unlikely(!mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1))) {
755 if (sq->db.doorbell) {
756 /* SQ is full, ring doorbell */
757 mlx5e_xmit_xdp_doorbell(sq);
758 sq->db.doorbell = false;
760 rq->stats.xdp_tx_full++;
764 dma_sync_single_for_device(sq->pdev, dma_addr, dma_len, PCI_DMA_TODEVICE);
768 dseg = (struct mlx5_wqe_data_seg *)eseg + 1;
770 /* copy the inline part if required */
771 if (sq->min_inline_mode != MLX5_INLINE_MODE_NONE) {
772 memcpy(eseg->inline_hdr.start, xdp->data, MLX5E_XDP_MIN_INLINE);
773 eseg->inline_hdr.sz = cpu_to_be16(MLX5E_XDP_MIN_INLINE);
774 dma_len -= MLX5E_XDP_MIN_INLINE;
775 dma_addr += MLX5E_XDP_MIN_INLINE;
779 /* write the dma part */
780 dseg->addr = cpu_to_be64(dma_addr);
781 dseg->byte_count = cpu_to_be32(dma_len);
783 cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8) | MLX5_OPCODE_SEND);
785 /* move page to reference to sq responsibility,
786 * and mark so it's not put back in page-cache.
788 rq->wqe.xdp_xmit = true;
792 sq->db.doorbell = true;
798 /* returns true if packet was consumed by xdp */
799 static inline int mlx5e_xdp_handle(struct mlx5e_rq *rq,
800 struct mlx5e_dma_info *di,
801 void *va, u16 *rx_headroom, u32 *len)
803 const struct bpf_prog *prog = READ_ONCE(rq->xdp_prog);
810 xdp.data = va + *rx_headroom;
811 xdp_set_data_meta_invalid(&xdp);
812 xdp.data_end = xdp.data + *len;
813 xdp.data_hard_start = va;
814 xdp.rxq = &rq->xdp_rxq;
816 act = bpf_prog_run_xdp(prog, &xdp);
819 *rx_headroom = xdp.data - xdp.data_hard_start;
820 *len = xdp.data_end - xdp.data;
823 if (unlikely(!mlx5e_xmit_xdp_frame(rq, di, &xdp)))
824 trace_xdp_exception(rq->netdev, prog, act);
827 bpf_warn_invalid_xdp_action(act);
829 trace_xdp_exception(rq->netdev, prog, act);
831 rq->stats.xdp_drop++;
837 struct sk_buff *skb_from_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe,
838 struct mlx5e_wqe_frag_info *wi, u32 cqe_bcnt)
840 struct mlx5e_dma_info *di = &wi->di;
841 u16 rx_headroom = rq->buff.headroom;
847 va = page_address(di->page) + wi->offset;
848 data = va + rx_headroom;
849 frag_size = MLX5_SKB_FRAG_SZ(rx_headroom + cqe_bcnt);
851 dma_sync_single_range_for_cpu(rq->pdev,
852 di->addr + wi->offset,
856 wi->offset += frag_size;
858 if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
864 consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt);
867 return NULL; /* page/packet was consumed by XDP */
869 skb = build_skb(va, frag_size);
870 if (unlikely(!skb)) {
871 rq->stats.buff_alloc_err++;
875 /* queue up for recycling/reuse */
876 page_ref_inc(di->page);
878 skb_reserve(skb, rx_headroom);
879 skb_put(skb, cqe_bcnt);
884 void mlx5e_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
886 struct mlx5e_wqe_frag_info *wi;
887 struct mlx5e_rx_wqe *wqe;
888 __be16 wqe_counter_be;
893 wqe_counter_be = cqe->wqe_counter;
894 wqe_counter = be16_to_cpu(wqe_counter_be);
895 wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
896 wi = &rq->wqe.frag_info[wqe_counter];
897 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
899 skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
901 /* probably for XDP */
902 if (rq->wqe.xdp_xmit) {
904 rq->wqe.xdp_xmit = false;
905 /* do not return page to cache, it will be returned on XDP_TX completion */
908 /* probably an XDP_DROP, save the page-reuse checks */
909 mlx5e_free_rx_wqe(rq, wi);
913 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
914 napi_gro_receive(rq->cq.napi, skb);
916 mlx5e_free_rx_wqe_reuse(rq, wi);
918 mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
919 &wqe->next.next_wqe_index);
922 #ifdef CONFIG_MLX5_ESWITCH
923 void mlx5e_handle_rx_cqe_rep(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
925 struct net_device *netdev = rq->netdev;
926 struct mlx5e_priv *priv = netdev_priv(netdev);
927 struct mlx5e_rep_priv *rpriv = priv->ppriv;
928 struct mlx5_eswitch_rep *rep = rpriv->rep;
929 struct mlx5e_wqe_frag_info *wi;
930 struct mlx5e_rx_wqe *wqe;
932 __be16 wqe_counter_be;
936 wqe_counter_be = cqe->wqe_counter;
937 wqe_counter = be16_to_cpu(wqe_counter_be);
938 wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
939 wi = &rq->wqe.frag_info[wqe_counter];
940 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
942 skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
944 if (rq->wqe.xdp_xmit) {
946 rq->wqe.xdp_xmit = false;
947 /* do not return page to cache, it will be returned on XDP_TX completion */
950 /* probably an XDP_DROP, save the page-reuse checks */
951 mlx5e_free_rx_wqe(rq, wi);
955 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
957 if (rep->vlan && skb_vlan_tag_present(skb))
960 napi_gro_receive(rq->cq.napi, skb);
962 mlx5e_free_rx_wqe_reuse(rq, wi);
964 mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
965 &wqe->next.next_wqe_index);
969 static inline void mlx5e_mpwqe_fill_rx_skb(struct mlx5e_rq *rq,
970 struct mlx5_cqe64 *cqe,
971 struct mlx5e_mpw_info *wi,
975 u16 stride_ix = mpwrq_get_cqe_stride_index(cqe);
976 u32 wqe_offset = stride_ix << rq->mpwqe.log_stride_sz;
977 u32 head_offset = wqe_offset & (PAGE_SIZE - 1);
978 u32 page_idx = wqe_offset >> PAGE_SHIFT;
979 u32 head_page_idx = page_idx;
980 u16 headlen = min_t(u16, MLX5_MPWRQ_SMALL_PACKET_THRESHOLD, cqe_bcnt);
981 u32 frag_offset = head_offset + headlen;
982 u16 byte_cnt = cqe_bcnt - headlen;
984 if (unlikely(frag_offset >= PAGE_SIZE)) {
986 frag_offset -= PAGE_SIZE;
990 u32 pg_consumed_bytes =
991 min_t(u32, PAGE_SIZE - frag_offset, byte_cnt);
993 mlx5e_add_skb_frag_mpwqe(rq, skb, wi, page_idx, frag_offset,
995 byte_cnt -= pg_consumed_bytes;
1000 mlx5e_copy_skb_header_mpwqe(rq->pdev, skb, wi, head_page_idx,
1001 head_offset, headlen);
1002 /* skb linear part was allocated with headlen and aligned to long */
1003 skb->tail += headlen;
1004 skb->len += headlen;
1007 void mlx5e_handle_rx_cqe_mpwrq(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1009 u16 cstrides = mpwrq_get_cqe_consumed_strides(cqe);
1010 u16 wqe_id = be16_to_cpu(cqe->wqe_id);
1011 struct mlx5e_mpw_info *wi = &rq->mpwqe.info[wqe_id];
1012 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_id);
1013 struct sk_buff *skb;
1016 wi->consumed_strides += cstrides;
1018 if (unlikely((cqe->op_own >> 4) != MLX5_CQE_RESP_SEND)) {
1019 rq->stats.wqe_err++;
1023 if (unlikely(mpwrq_is_filler_cqe(cqe))) {
1024 rq->stats.mpwqe_filler++;
1028 skb = napi_alloc_skb(rq->cq.napi,
1029 ALIGN(MLX5_MPWRQ_SMALL_PACKET_THRESHOLD,
1031 if (unlikely(!skb)) {
1032 rq->stats.buff_alloc_err++;
1036 prefetchw(skb->data);
1037 cqe_bcnt = mpwrq_get_cqe_byte_cnt(cqe);
1039 mlx5e_mpwqe_fill_rx_skb(rq, cqe, wi, cqe_bcnt, skb);
1040 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1041 napi_gro_receive(rq->cq.napi, skb);
1044 if (likely(wi->consumed_strides < rq->mpwqe.num_strides))
1047 mlx5e_free_rx_mpwqe(rq, wi);
1048 mlx5_wq_ll_pop(&rq->wq, cqe->wqe_id, &wqe->next.next_wqe_index);
1051 int mlx5e_poll_rx_cq(struct mlx5e_cq *cq, int budget)
1053 struct mlx5e_rq *rq = container_of(cq, struct mlx5e_rq, cq);
1054 struct mlx5e_xdpsq *xdpsq;
1055 struct mlx5_cqe64 *cqe;
1058 if (unlikely(!MLX5E_TEST_BIT(rq->state, MLX5E_RQ_STATE_ENABLED)))
1061 if (cq->decmprs_left)
1062 work_done += mlx5e_decompress_cqes_cont(rq, cq, 0, budget);
1064 cqe = mlx5_cqwq_get_cqe(&cq->wq);
1071 if (mlx5_get_cqe_format(cqe) == MLX5_COMPRESSED) {
1073 mlx5e_decompress_cqes_start(rq, cq,
1074 budget - work_done);
1078 mlx5_cqwq_pop(&cq->wq);
1080 rq->handle_rx_cqe(rq, cqe);
1081 } while ((++work_done < budget) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1083 if (xdpsq->db.doorbell) {
1084 mlx5e_xmit_xdp_doorbell(xdpsq);
1085 xdpsq->db.doorbell = false;
1088 mlx5_cqwq_update_db_record(&cq->wq);
1090 /* ensure cq space is freed before enabling more cqes */
1096 bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
1098 struct mlx5e_xdpsq *sq;
1099 struct mlx5_cqe64 *cqe;
1100 struct mlx5e_rq *rq;
1104 sq = container_of(cq, struct mlx5e_xdpsq, cq);
1106 if (unlikely(!MLX5E_TEST_BIT(sq->state, MLX5E_SQ_STATE_ENABLED)))
1109 cqe = mlx5_cqwq_get_cqe(&cq->wq);
1113 rq = container_of(sq, struct mlx5e_rq, xdpsq);
1115 /* sq->cc must be updated only after mlx5_cqwq_update_db_record(),
1116 * otherwise a cq overrun may occur
1125 mlx5_cqwq_pop(&cq->wq);
1127 wqe_counter = be16_to_cpu(cqe->wqe_counter);
1130 struct mlx5e_dma_info *di;
1133 last_wqe = (sqcc == wqe_counter);
1135 ci = sqcc & sq->wq.sz_m1;
1136 di = &sq->db.di[ci];
1139 /* Recycle RX page */
1140 mlx5e_page_release(rq, di, true);
1141 } while (!last_wqe);
1142 } while ((++i < MLX5E_TX_CQ_POLL_BUDGET) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
1144 mlx5_cqwq_update_db_record(&cq->wq);
1146 /* ensure cq space is freed before enabling more cqes */
1150 return (i == MLX5E_TX_CQ_POLL_BUDGET);
1153 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
1155 struct mlx5e_rq *rq = container_of(sq, struct mlx5e_rq, xdpsq);
1156 struct mlx5e_dma_info *di;
1159 while (sq->cc != sq->pc) {
1160 ci = sq->cc & sq->wq.sz_m1;
1161 di = &sq->db.di[ci];
1164 mlx5e_page_release(rq, di, false);
1168 #ifdef CONFIG_MLX5_CORE_IPOIB
1170 #define MLX5_IB_GRH_DGID_OFFSET 24
1171 #define MLX5_GID_SIZE 16
1173 static inline void mlx5i_complete_rx_cqe(struct mlx5e_rq *rq,
1174 struct mlx5_cqe64 *cqe,
1176 struct sk_buff *skb)
1178 struct hwtstamp_config *tstamp;
1179 struct net_device *netdev;
1180 struct mlx5e_priv *priv;
1181 char *pseudo_header;
1186 qpn = be32_to_cpu(cqe->sop_drop_qpn) & 0xffffff;
1187 netdev = mlx5i_pkey_get_netdev(rq->netdev, qpn);
1189 /* No mapping present, cannot process SKB. This might happen if a child
1190 * interface is going down while having unprocessed CQEs on parent RQ
1192 if (unlikely(!netdev)) {
1193 /* TODO: add drop counters support */
1195 pr_warn_once("Unable to map QPN %u to dev - dropping skb\n", qpn);
1199 priv = mlx5i_epriv(netdev);
1200 tstamp = &priv->tstamp;
1202 g = (be32_to_cpu(cqe->flags_rqpn) >> 28) & 3;
1203 dgid = skb->data + MLX5_IB_GRH_DGID_OFFSET;
1204 if ((!g) || dgid[0] != 0xff)
1205 skb->pkt_type = PACKET_HOST;
1206 else if (memcmp(dgid, netdev->broadcast + 4, MLX5_GID_SIZE) == 0)
1207 skb->pkt_type = PACKET_BROADCAST;
1209 skb->pkt_type = PACKET_MULTICAST;
1211 /* TODO: IB/ipoib: Allow mcast packets from other VFs
1212 * 68996a6e760e5c74654723eeb57bf65628ae87f4
1215 skb_pull(skb, MLX5_IB_GRH_BYTES);
1217 skb->protocol = *((__be16 *)(skb->data));
1219 skb->ip_summed = CHECKSUM_COMPLETE;
1220 skb->csum = csum_unfold((__force __sum16)cqe->check_sum);
1222 if (unlikely(mlx5e_rx_hw_stamp(tstamp)))
1223 skb_hwtstamps(skb)->hwtstamp =
1224 mlx5_timecounter_cyc2time(rq->clock, get_cqe_ts(cqe));
1226 skb_record_rx_queue(skb, rq->ix);
1228 if (likely(netdev->features & NETIF_F_RXHASH))
1229 mlx5e_skb_set_hash(cqe, skb);
1231 /* 20 bytes of ipoib header and 4 for encap existing */
1232 pseudo_header = skb_push(skb, MLX5_IPOIB_PSEUDO_LEN);
1233 memset(pseudo_header, 0, MLX5_IPOIB_PSEUDO_LEN);
1234 skb_reset_mac_header(skb);
1235 skb_pull(skb, MLX5_IPOIB_HARD_LEN);
1239 rq->stats.csum_complete++;
1240 rq->stats.packets++;
1241 rq->stats.bytes += cqe_bcnt;
1244 void mlx5i_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1246 struct mlx5e_wqe_frag_info *wi;
1247 struct mlx5e_rx_wqe *wqe;
1248 __be16 wqe_counter_be;
1249 struct sk_buff *skb;
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);
1259 skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1263 mlx5i_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1264 if (unlikely(!skb->dev)) {
1265 dev_kfree_skb_any(skb);
1268 napi_gro_receive(rq->cq.napi, skb);
1271 mlx5e_free_rx_wqe_reuse(rq, wi);
1272 mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1273 &wqe->next.next_wqe_index);
1276 #endif /* CONFIG_MLX5_CORE_IPOIB */
1278 #ifdef CONFIG_MLX5_EN_IPSEC
1280 void mlx5e_ipsec_handle_rx_cqe(struct mlx5e_rq *rq, struct mlx5_cqe64 *cqe)
1282 struct mlx5e_wqe_frag_info *wi;
1283 struct mlx5e_rx_wqe *wqe;
1284 __be16 wqe_counter_be;
1285 struct sk_buff *skb;
1289 wqe_counter_be = cqe->wqe_counter;
1290 wqe_counter = be16_to_cpu(wqe_counter_be);
1291 wqe = mlx5_wq_ll_get_wqe(&rq->wq, wqe_counter);
1292 wi = &rq->wqe.frag_info[wqe_counter];
1293 cqe_bcnt = be32_to_cpu(cqe->byte_cnt);
1295 skb = skb_from_cqe(rq, cqe, wi, cqe_bcnt);
1296 if (unlikely(!skb)) {
1297 /* a DROP, save the page-reuse checks */
1298 mlx5e_free_rx_wqe(rq, wi);
1301 skb = mlx5e_ipsec_handle_rx_skb(rq->netdev, skb);
1302 if (unlikely(!skb)) {
1303 mlx5e_free_rx_wqe(rq, wi);
1307 mlx5e_complete_rx_cqe(rq, cqe, cqe_bcnt, skb);
1308 napi_gro_receive(rq->cq.napi, skb);
1310 mlx5e_free_rx_wqe_reuse(rq, wi);
1312 mlx5_wq_ll_pop(&rq->wq, wqe_counter_be,
1313 &wqe->next.next_wqe_index);
1316 #endif /* CONFIG_MLX5_EN_IPSEC */