1 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
2 /* Copyright (C) 2015-2019 Netronome Systems, Inc. */
4 #include <linux/bpf_trace.h>
5 #include <linux/netdevice.h>
7 #include "../nfp_app.h"
8 #include "../nfp_net.h"
9 #include "../nfp_net_dp.h"
10 #include "../nfp_net_xsk.h"
11 #include "../crypto/crypto.h"
12 #include "../crypto/fw.h"
15 /* Transmit processing
17 * One queue controller peripheral queue is used for transmit. The
18 * driver en-queues packets for transmit by advancing the write
19 * pointer. The device indicates that packets have transmitted by
20 * advancing the read pointer. The driver maintains a local copy of
21 * the read and write pointer in @struct nfp_net_tx_ring. The driver
22 * keeps @wr_p in sync with the queue controller write pointer and can
23 * determine how many packets have been transmitted by comparing its
24 * copy of the read pointer @rd_p with the read pointer maintained by
25 * the queue controller peripheral.
28 /* Wrappers for deciding when to stop and restart TX queues */
29 static int nfp_nfd3_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
31 return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4);
34 static int nfp_nfd3_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
36 return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1);
40 * nfp_nfd3_tx_ring_stop() - stop tx ring
42 * @tx_ring: driver tx queue structure
44 * Safely stop TX ring. Remember that while we are running .start_xmit()
45 * someone else may be cleaning the TX ring completions so we need to be
49 nfp_nfd3_tx_ring_stop(struct netdev_queue *nd_q,
50 struct nfp_net_tx_ring *tx_ring)
52 netif_tx_stop_queue(nd_q);
54 /* We can race with the TX completion out of NAPI so recheck */
56 if (unlikely(nfp_nfd3_tx_ring_should_wake(tx_ring)))
57 netif_tx_start_queue(nd_q);
61 * nfp_nfd3_tx_tso() - Set up Tx descriptor for LSO
62 * @r_vec: per-ring structure
63 * @txbuf: Pointer to driver soft TX descriptor
64 * @txd: Pointer to HW TX descriptor
65 * @skb: Pointer to SKB
66 * @md_bytes: Prepend length
68 * Set up Tx descriptor for LSO, do nothing for non-LSO skbs.
69 * Return error on packet header greater than maximum supported LSO header size.
72 nfp_nfd3_tx_tso(struct nfp_net_r_vector *r_vec, struct nfp_nfd3_tx_buf *txbuf,
73 struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb, u32 md_bytes)
75 u32 l3_offset, l4_offset, hdrlen;
81 if (!skb->encapsulation) {
82 l3_offset = skb_network_offset(skb);
83 l4_offset = skb_transport_offset(skb);
84 hdrlen = skb_transport_offset(skb) + tcp_hdrlen(skb);
86 l3_offset = skb_inner_network_offset(skb);
87 l4_offset = skb_inner_transport_offset(skb);
88 hdrlen = skb_inner_transport_header(skb) - skb->data +
89 inner_tcp_hdrlen(skb);
92 txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs;
93 txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1);
95 mss = skb_shinfo(skb)->gso_size & NFD3_DESC_TX_MSS_MASK;
96 txd->l3_offset = l3_offset - md_bytes;
97 txd->l4_offset = l4_offset - md_bytes;
98 txd->lso_hdrlen = hdrlen - md_bytes;
99 txd->mss = cpu_to_le16(mss);
100 txd->flags |= NFD3_DESC_TX_LSO;
102 u64_stats_update_begin(&r_vec->tx_sync);
104 u64_stats_update_end(&r_vec->tx_sync);
108 * nfp_nfd3_tx_csum() - Set TX CSUM offload flags in TX descriptor
109 * @dp: NFP Net data path struct
110 * @r_vec: per-ring structure
111 * @txbuf: Pointer to driver soft TX descriptor
112 * @txd: Pointer to TX descriptor
113 * @skb: Pointer to SKB
115 * This function sets the TX checksum flags in the TX descriptor based
116 * on the configuration and the protocol of the packet to be transmitted.
119 nfp_nfd3_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
120 struct nfp_nfd3_tx_buf *txbuf, struct nfp_nfd3_tx_desc *txd,
123 struct ipv6hdr *ipv6h;
127 if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
130 if (skb->ip_summed != CHECKSUM_PARTIAL)
133 txd->flags |= NFD3_DESC_TX_CSUM;
134 if (skb->encapsulation)
135 txd->flags |= NFD3_DESC_TX_ENCAP;
137 iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
138 ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
140 if (iph->version == 4) {
141 txd->flags |= NFD3_DESC_TX_IP4_CSUM;
142 l4_hdr = iph->protocol;
143 } else if (ipv6h->version == 6) {
144 l4_hdr = ipv6h->nexthdr;
146 nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
152 txd->flags |= NFD3_DESC_TX_TCP_CSUM;
155 txd->flags |= NFD3_DESC_TX_UDP_CSUM;
158 nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr);
162 u64_stats_update_begin(&r_vec->tx_sync);
163 if (skb->encapsulation)
164 r_vec->hw_csum_tx_inner += txbuf->pkt_cnt;
166 r_vec->hw_csum_tx += txbuf->pkt_cnt;
167 u64_stats_update_end(&r_vec->tx_sync);
170 static int nfp_nfd3_prep_tx_meta(struct sk_buff *skb, u64 tls_handle)
172 struct metadata_dst *md_dst = skb_metadata_dst(skb);
177 if (likely(!md_dst && !tls_handle))
179 if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX)) {
185 md_bytes = 4 + !!md_dst * 4 + !!tls_handle * 8;
187 if (unlikely(skb_cow_head(skb, md_bytes)))
191 data = skb_push(skb, md_bytes) + md_bytes;
194 put_unaligned_be32(md_dst->u.port_info.port_id, data);
195 meta_id = NFP_NET_META_PORTID;
198 /* conn handle is opaque, we just use u64 to be able to quickly
202 memcpy(data, &tls_handle, sizeof(tls_handle));
203 meta_id <<= NFP_NET_META_FIELD_SIZE;
204 meta_id |= NFP_NET_META_CONN_HANDLE;
208 put_unaligned_be32(meta_id, data);
214 * nfp_nfd3_tx() - Main transmit entry point
215 * @skb: SKB to transmit
216 * @netdev: netdev structure
218 * Return: NETDEV_TX_OK on success.
220 netdev_tx_t nfp_nfd3_tx(struct sk_buff *skb, struct net_device *netdev)
222 struct nfp_net *nn = netdev_priv(netdev);
223 int f, nr_frags, wr_idx, md_bytes;
224 struct nfp_net_tx_ring *tx_ring;
225 struct nfp_net_r_vector *r_vec;
226 struct nfp_nfd3_tx_buf *txbuf;
227 struct nfp_nfd3_tx_desc *txd;
228 struct netdev_queue *nd_q;
229 const skb_frag_t *frag;
230 struct nfp_net_dp *dp;
237 qidx = skb_get_queue_mapping(skb);
238 tx_ring = &dp->tx_rings[qidx];
239 r_vec = tx_ring->r_vec;
241 nr_frags = skb_shinfo(skb)->nr_frags;
243 if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) {
244 nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
245 qidx, tx_ring->wr_p, tx_ring->rd_p);
246 nd_q = netdev_get_tx_queue(dp->netdev, qidx);
247 netif_tx_stop_queue(nd_q);
248 nfp_net_tx_xmit_more_flush(tx_ring);
249 u64_stats_update_begin(&r_vec->tx_sync);
251 u64_stats_update_end(&r_vec->tx_sync);
252 return NETDEV_TX_BUSY;
255 skb = nfp_net_tls_tx(dp, r_vec, skb, &tls_handle, &nr_frags);
256 if (unlikely(!skb)) {
257 nfp_net_tx_xmit_more_flush(tx_ring);
261 md_bytes = nfp_nfd3_prep_tx_meta(skb, tls_handle);
262 if (unlikely(md_bytes < 0))
265 /* Start with the head skbuf */
266 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
268 if (dma_mapping_error(dp->dev, dma_addr))
271 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
273 /* Stash the soft descriptor of the head then initialize it */
274 txbuf = &tx_ring->txbufs[wr_idx];
276 txbuf->dma_addr = dma_addr;
279 txbuf->real_len = skb->len;
281 /* Build TX descriptor */
282 txd = &tx_ring->txds[wr_idx];
283 txd->offset_eop = (nr_frags ? 0 : NFD3_DESC_TX_EOP) | md_bytes;
284 txd->dma_len = cpu_to_le16(skb_headlen(skb));
285 nfp_desc_set_dma_addr(txd, dma_addr);
286 txd->data_len = cpu_to_le16(skb->len);
292 /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */
293 nfp_nfd3_tx_tso(r_vec, txbuf, txd, skb, md_bytes);
294 nfp_nfd3_tx_csum(dp, r_vec, txbuf, txd, skb);
295 if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) {
296 txd->flags |= NFD3_DESC_TX_VLAN;
297 txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
304 /* all descs must match except for in addr, length and eop */
305 second_half = txd->vals8[1];
307 for (f = 0; f < nr_frags; f++) {
308 frag = &skb_shinfo(skb)->frags[f];
309 fsize = skb_frag_size(frag);
311 dma_addr = skb_frag_dma_map(dp->dev, frag, 0,
312 fsize, DMA_TO_DEVICE);
313 if (dma_mapping_error(dp->dev, dma_addr))
316 wr_idx = D_IDX(tx_ring, wr_idx + 1);
317 tx_ring->txbufs[wr_idx].skb = skb;
318 tx_ring->txbufs[wr_idx].dma_addr = dma_addr;
319 tx_ring->txbufs[wr_idx].fidx = f;
321 txd = &tx_ring->txds[wr_idx];
322 txd->dma_len = cpu_to_le16(fsize);
323 nfp_desc_set_dma_addr(txd, dma_addr);
324 txd->offset_eop = md_bytes |
325 ((f == nr_frags - 1) ? NFD3_DESC_TX_EOP : 0);
326 txd->vals8[1] = second_half;
329 u64_stats_update_begin(&r_vec->tx_sync);
331 u64_stats_update_end(&r_vec->tx_sync);
334 skb_tx_timestamp(skb);
336 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
338 tx_ring->wr_p += nr_frags + 1;
339 if (nfp_nfd3_tx_ring_should_stop(tx_ring))
340 nfp_nfd3_tx_ring_stop(nd_q, tx_ring);
342 tx_ring->wr_ptr_add += nr_frags + 1;
343 if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, netdev_xmit_more()))
344 nfp_net_tx_xmit_more_flush(tx_ring);
350 frag = &skb_shinfo(skb)->frags[f];
351 dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
352 skb_frag_size(frag), DMA_TO_DEVICE);
353 tx_ring->txbufs[wr_idx].skb = NULL;
354 tx_ring->txbufs[wr_idx].dma_addr = 0;
355 tx_ring->txbufs[wr_idx].fidx = -2;
358 wr_idx += tx_ring->cnt;
360 dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
361 skb_headlen(skb), DMA_TO_DEVICE);
362 tx_ring->txbufs[wr_idx].skb = NULL;
363 tx_ring->txbufs[wr_idx].dma_addr = 0;
364 tx_ring->txbufs[wr_idx].fidx = -2;
366 nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
368 nfp_net_tx_xmit_more_flush(tx_ring);
369 u64_stats_update_begin(&r_vec->tx_sync);
371 u64_stats_update_end(&r_vec->tx_sync);
372 nfp_net_tls_tx_undo(skb, tls_handle);
373 dev_kfree_skb_any(skb);
378 * nfp_nfd3_tx_complete() - Handled completed TX packets
379 * @tx_ring: TX ring structure
380 * @budget: NAPI budget (only used as bool to determine if in NAPI context)
382 void nfp_nfd3_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
384 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
385 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
386 u32 done_pkts = 0, done_bytes = 0;
387 struct netdev_queue *nd_q;
391 if (tx_ring->wr_p == tx_ring->rd_p)
394 /* Work out how many descriptors have been transmitted */
395 qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
397 if (qcp_rd_p == tx_ring->qcp_rd_p)
400 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
403 const skb_frag_t *frag;
404 struct nfp_nfd3_tx_buf *tx_buf;
409 idx = D_IDX(tx_ring, tx_ring->rd_p++);
410 tx_buf = &tx_ring->txbufs[idx];
416 nr_frags = skb_shinfo(skb)->nr_frags;
421 dma_unmap_single(dp->dev, tx_buf->dma_addr,
422 skb_headlen(skb), DMA_TO_DEVICE);
424 done_pkts += tx_buf->pkt_cnt;
425 done_bytes += tx_buf->real_len;
428 frag = &skb_shinfo(skb)->frags[fidx];
429 dma_unmap_page(dp->dev, tx_buf->dma_addr,
430 skb_frag_size(frag), DMA_TO_DEVICE);
433 /* check for last gather fragment */
434 if (fidx == nr_frags - 1)
435 napi_consume_skb(skb, budget);
437 tx_buf->dma_addr = 0;
442 tx_ring->qcp_rd_p = qcp_rd_p;
444 u64_stats_update_begin(&r_vec->tx_sync);
445 r_vec->tx_bytes += done_bytes;
446 r_vec->tx_pkts += done_pkts;
447 u64_stats_update_end(&r_vec->tx_sync);
452 nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
453 netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
454 if (nfp_nfd3_tx_ring_should_wake(tx_ring)) {
455 /* Make sure TX thread will see updated tx_ring->rd_p */
458 if (unlikely(netif_tx_queue_stopped(nd_q)))
459 netif_tx_wake_queue(nd_q);
462 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
463 "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
464 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
467 static bool nfp_nfd3_xdp_complete(struct nfp_net_tx_ring *tx_ring)
469 struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
470 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
471 u32 done_pkts = 0, done_bytes = 0;
476 /* Work out how many descriptors have been transmitted */
477 qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
479 if (qcp_rd_p == tx_ring->qcp_rd_p)
482 todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
484 done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
485 todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
487 tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
491 idx = D_IDX(tx_ring, tx_ring->rd_p);
494 done_bytes += tx_ring->txbufs[idx].real_len;
497 u64_stats_update_begin(&r_vec->tx_sync);
498 r_vec->tx_bytes += done_bytes;
499 r_vec->tx_pkts += done_pkts;
500 u64_stats_update_end(&r_vec->tx_sync);
502 WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
503 "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
504 tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
509 /* Receive processing
513 nfp_nfd3_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
518 frag = napi_alloc_frag(dp->fl_bufsz);
524 page = dev_alloc_page();
527 frag = page_address(page);
530 *dma_addr = nfp_net_dma_map_rx(dp, frag);
531 if (dma_mapping_error(dp->dev, *dma_addr)) {
532 nfp_net_free_frag(frag, dp->xdp_prog);
533 nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
541 * nfp_nfd3_rx_give_one() - Put mapped skb on the software and hardware rings
542 * @dp: NFP Net data path struct
543 * @rx_ring: RX ring structure
544 * @frag: page fragment buffer
545 * @dma_addr: DMA address of skb mapping
548 nfp_nfd3_rx_give_one(const struct nfp_net_dp *dp,
549 struct nfp_net_rx_ring *rx_ring,
550 void *frag, dma_addr_t dma_addr)
554 wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
556 nfp_net_dma_sync_dev_rx(dp, dma_addr);
558 /* Stash SKB and DMA address away */
559 rx_ring->rxbufs[wr_idx].frag = frag;
560 rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
562 /* Fill freelist descriptor */
563 rx_ring->rxds[wr_idx].fld.reserved = 0;
564 rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
565 nfp_desc_set_dma_addr(&rx_ring->rxds[wr_idx].fld,
566 dma_addr + dp->rx_dma_off);
569 if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
570 /* Update write pointer of the freelist queue. Make
571 * sure all writes are flushed before telling the hardware.
574 nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
579 * nfp_nfd3_rx_ring_fill_freelist() - Give buffers from the ring to FW
580 * @dp: NFP Net data path struct
581 * @rx_ring: RX ring to fill
583 void nfp_nfd3_rx_ring_fill_freelist(struct nfp_net_dp *dp,
584 struct nfp_net_rx_ring *rx_ring)
588 if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
589 return nfp_net_xsk_rx_ring_fill_freelist(rx_ring);
591 for (i = 0; i < rx_ring->cnt - 1; i++)
592 nfp_nfd3_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
593 rx_ring->rxbufs[i].dma_addr);
597 * nfp_nfd3_rx_csum_has_errors() - group check if rxd has any csum errors
598 * @flags: RX descriptor flags field in CPU byte order
600 static int nfp_nfd3_rx_csum_has_errors(u16 flags)
602 u16 csum_all_checked, csum_all_ok;
604 csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
605 csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
607 return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
611 * nfp_nfd3_rx_csum() - set SKB checksum field based on RX descriptor flags
612 * @dp: NFP Net data path struct
613 * @r_vec: per-ring structure
614 * @rxd: Pointer to RX descriptor
615 * @meta: Parsed metadata prepend
616 * @skb: Pointer to SKB
619 nfp_nfd3_rx_csum(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
620 const struct nfp_net_rx_desc *rxd,
621 const struct nfp_meta_parsed *meta, struct sk_buff *skb)
623 skb_checksum_none_assert(skb);
625 if (!(dp->netdev->features & NETIF_F_RXCSUM))
628 if (meta->csum_type) {
629 skb->ip_summed = meta->csum_type;
630 skb->csum = meta->csum;
631 u64_stats_update_begin(&r_vec->rx_sync);
632 r_vec->hw_csum_rx_complete++;
633 u64_stats_update_end(&r_vec->rx_sync);
637 if (nfp_nfd3_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
638 u64_stats_update_begin(&r_vec->rx_sync);
639 r_vec->hw_csum_rx_error++;
640 u64_stats_update_end(&r_vec->rx_sync);
644 /* Assume that the firmware will never report inner CSUM_OK unless outer
645 * L4 headers were successfully parsed. FW will always report zero UDP
646 * checksum as CSUM_OK.
648 if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
649 rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
650 __skb_incr_checksum_unnecessary(skb);
651 u64_stats_update_begin(&r_vec->rx_sync);
652 r_vec->hw_csum_rx_ok++;
653 u64_stats_update_end(&r_vec->rx_sync);
656 if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
657 rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
658 __skb_incr_checksum_unnecessary(skb);
659 u64_stats_update_begin(&r_vec->rx_sync);
660 r_vec->hw_csum_rx_inner_ok++;
661 u64_stats_update_end(&r_vec->rx_sync);
666 nfp_nfd3_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
667 unsigned int type, __be32 *hash)
669 if (!(netdev->features & NETIF_F_RXHASH))
673 case NFP_NET_RSS_IPV4:
674 case NFP_NET_RSS_IPV6:
675 case NFP_NET_RSS_IPV6_EX:
676 meta->hash_type = PKT_HASH_TYPE_L3;
679 meta->hash_type = PKT_HASH_TYPE_L4;
683 meta->hash = get_unaligned_be32(hash);
687 nfp_nfd3_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
688 void *data, struct nfp_net_rx_desc *rxd)
690 struct nfp_net_rx_hash *rx_hash = data;
692 if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
695 nfp_nfd3_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type),
700 nfp_nfd3_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
701 void *data, void *pkt, unsigned int pkt_len, int meta_len)
705 meta_info = get_unaligned_be32(data);
709 switch (meta_info & NFP_NET_META_FIELD_MASK) {
710 case NFP_NET_META_HASH:
711 meta_info >>= NFP_NET_META_FIELD_SIZE;
712 nfp_nfd3_set_hash(netdev, meta,
713 meta_info & NFP_NET_META_FIELD_MASK,
717 case NFP_NET_META_MARK:
718 meta->mark = get_unaligned_be32(data);
721 case NFP_NET_META_PORTID:
722 meta->portid = get_unaligned_be32(data);
725 case NFP_NET_META_CSUM:
726 meta->csum_type = CHECKSUM_COMPLETE;
728 (__force __wsum)__get_unaligned_cpu32(data);
731 case NFP_NET_META_RESYNC_INFO:
732 if (nfp_net_tls_rx_resync_req(netdev, data, pkt,
735 data += sizeof(struct nfp_net_tls_resync_req);
741 meta_info >>= NFP_NET_META_FIELD_SIZE;
748 nfp_nfd3_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
749 struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf,
752 u64_stats_update_begin(&r_vec->rx_sync);
754 /* If we have both skb and rxbuf the replacement buffer allocation
755 * must have failed, count this as an alloc failure.
758 r_vec->rx_replace_buf_alloc_fail++;
759 u64_stats_update_end(&r_vec->rx_sync);
761 /* skb is build based on the frag, free_skb() would free the frag
762 * so to be able to reuse it we need an extra ref.
764 if (skb && rxbuf && skb->head == rxbuf->frag)
765 page_ref_inc(virt_to_head_page(rxbuf->frag));
767 nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
769 dev_kfree_skb_any(skb);
773 nfp_nfd3_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
774 struct nfp_net_tx_ring *tx_ring,
775 struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
776 unsigned int pkt_len, bool *completed)
778 unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA;
779 struct nfp_nfd3_tx_buf *txbuf;
780 struct nfp_nfd3_tx_desc *txd;
783 /* Reject if xdp_adjust_tail grow packet beyond DMA area */
784 if (pkt_len + dma_off > dma_map_sz)
787 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
789 nfp_nfd3_xdp_complete(tx_ring);
793 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
794 nfp_nfd3_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
800 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
802 /* Stash the soft descriptor of the head then initialize it */
803 txbuf = &tx_ring->txbufs[wr_idx];
805 nfp_nfd3_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr);
807 txbuf->frag = rxbuf->frag;
808 txbuf->dma_addr = rxbuf->dma_addr;
811 txbuf->real_len = pkt_len;
813 dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
814 pkt_len, DMA_BIDIRECTIONAL);
816 /* Build TX descriptor */
817 txd = &tx_ring->txds[wr_idx];
818 txd->offset_eop = NFD3_DESC_TX_EOP;
819 txd->dma_len = cpu_to_le16(pkt_len);
820 nfp_desc_set_dma_addr(txd, rxbuf->dma_addr + dma_off);
821 txd->data_len = cpu_to_le16(pkt_len);
828 tx_ring->wr_ptr_add++;
833 * nfp_nfd3_rx() - receive up to @budget packets on @rx_ring
834 * @rx_ring: RX ring to receive from
835 * @budget: NAPI budget
837 * Note, this function is separated out from the napi poll function to
838 * more cleanly separate packet receive code from other bookkeeping
839 * functions performed in the napi poll function.
841 * Return: Number of packets received.
843 static int nfp_nfd3_rx(struct nfp_net_rx_ring *rx_ring, int budget)
845 struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
846 struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
847 struct nfp_net_tx_ring *tx_ring;
848 struct bpf_prog *xdp_prog;
849 bool xdp_tx_cmpl = false;
850 unsigned int true_bufsz;
856 xdp_prog = READ_ONCE(dp->xdp_prog);
857 true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
858 xdp_init_buff(&xdp, PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM,
860 tx_ring = r_vec->xdp_ring;
862 while (pkts_polled < budget) {
863 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
864 struct nfp_net_rx_buf *rxbuf;
865 struct nfp_net_rx_desc *rxd;
866 struct nfp_meta_parsed meta;
867 bool redir_egress = false;
868 struct net_device *netdev;
869 dma_addr_t new_dma_addr;
870 u32 meta_len_xdp = 0;
873 idx = D_IDX(rx_ring, rx_ring->rd_p);
875 rxd = &rx_ring->rxds[idx];
876 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
879 /* Memory barrier to ensure that we won't do other reads
884 memset(&meta, 0, sizeof(meta));
889 rxbuf = &rx_ring->rxbufs[idx];
891 * <-- [rx_offset] -->
892 * ---------------------------------------------------------
893 * | [XX] | metadata | packet | XXXX |
894 * ---------------------------------------------------------
895 * <---------------- data_len --------------->
897 * The rx_offset is fixed for all packets, the meta_len can vary
898 * on a packet by packet basis. If rx_offset is set to zero
899 * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
900 * buffer and is immediately followed by the packet (no [XX]).
902 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
903 data_len = le16_to_cpu(rxd->rxd.data_len);
904 pkt_len = data_len - meta_len;
906 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
907 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
910 pkt_off += dp->rx_offset;
911 meta_off = pkt_off - meta_len;
914 u64_stats_update_begin(&r_vec->rx_sync);
916 r_vec->rx_bytes += pkt_len;
917 u64_stats_update_end(&r_vec->rx_sync);
919 if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
920 (dp->rx_offset && meta_len > dp->rx_offset))) {
921 nn_dp_warn(dp, "oversized RX packet metadata %u\n",
923 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
927 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
930 if (!dp->chained_metadata_format) {
931 nfp_nfd3_set_hash_desc(dp->netdev, &meta,
932 rxbuf->frag + meta_off, rxd);
933 } else if (meta_len) {
934 if (unlikely(nfp_nfd3_parse_meta(dp->netdev, &meta,
935 rxbuf->frag + meta_off,
936 rxbuf->frag + pkt_off,
937 pkt_len, meta_len))) {
938 nn_dp_warn(dp, "invalid RX packet metadata\n");
939 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
945 if (xdp_prog && !meta.portid) {
946 void *orig_data = rxbuf->frag + pkt_off;
947 unsigned int dma_off;
950 xdp_prepare_buff(&xdp,
951 rxbuf->frag + NFP_NET_RX_BUF_HEADROOM,
952 pkt_off - NFP_NET_RX_BUF_HEADROOM,
955 act = bpf_prog_run_xdp(xdp_prog, &xdp);
957 pkt_len = xdp.data_end - xdp.data;
958 pkt_off += xdp.data - orig_data;
962 meta_len_xdp = xdp.data - xdp.data_meta;
965 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
966 if (unlikely(!nfp_nfd3_tx_xdp_buf(dp, rx_ring,
972 trace_xdp_exception(dp->netdev,
976 bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
979 trace_xdp_exception(dp->netdev, xdp_prog, act);
982 nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
988 if (likely(!meta.portid)) {
990 } else if (meta.portid == NFP_META_PORT_ID_CTRL) {
991 struct nfp_net *nn = netdev_priv(dp->netdev);
993 nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
995 nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
1001 nn = netdev_priv(dp->netdev);
1002 netdev = nfp_app_dev_get(nn->app, meta.portid,
1004 if (unlikely(!netdev)) {
1005 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
1010 if (nfp_netdev_is_nfp_repr(netdev))
1011 nfp_repr_inc_rx_stats(netdev, pkt_len);
1014 skb = build_skb(rxbuf->frag, true_bufsz);
1015 if (unlikely(!skb)) {
1016 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1019 new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1020 if (unlikely(!new_frag)) {
1021 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1025 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1027 nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1029 skb_reserve(skb, pkt_off);
1030 skb_put(skb, pkt_len);
1032 skb->mark = meta.mark;
1033 skb_set_hash(skb, meta.hash, meta.hash_type);
1035 skb_record_rx_queue(skb, rx_ring->idx);
1036 skb->protocol = eth_type_trans(skb, netdev);
1038 nfp_nfd3_rx_csum(dp, r_vec, rxd, &meta, skb);
1040 #ifdef CONFIG_TLS_DEVICE
1041 if (rxd->rxd.flags & PCIE_DESC_RX_DECRYPTED) {
1042 skb->decrypted = true;
1043 u64_stats_update_begin(&r_vec->rx_sync);
1045 u64_stats_update_end(&r_vec->rx_sync);
1049 if (rxd->rxd.flags & PCIE_DESC_RX_VLAN)
1050 __vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q),
1051 le16_to_cpu(rxd->rxd.vlan));
1053 skb_metadata_set(skb, meta_len_xdp);
1055 if (likely(!redir_egress)) {
1056 napi_gro_receive(&rx_ring->r_vec->napi, skb);
1059 skb_reset_network_header(skb);
1060 __skb_push(skb, ETH_HLEN);
1061 dev_queue_xmit(skb);
1066 if (tx_ring->wr_ptr_add)
1067 nfp_net_tx_xmit_more_flush(tx_ring);
1068 else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1070 if (!nfp_nfd3_xdp_complete(tx_ring))
1071 pkts_polled = budget;
1078 * nfp_nfd3_poll() - napi poll function
1079 * @napi: NAPI structure
1080 * @budget: NAPI budget
1082 * Return: number of packets polled.
1084 int nfp_nfd3_poll(struct napi_struct *napi, int budget)
1086 struct nfp_net_r_vector *r_vec =
1087 container_of(napi, struct nfp_net_r_vector, napi);
1088 unsigned int pkts_polled = 0;
1091 nfp_nfd3_tx_complete(r_vec->tx_ring, budget);
1093 pkts_polled = nfp_nfd3_rx(r_vec->rx_ring, budget);
1095 if (pkts_polled < budget)
1096 if (napi_complete_done(napi, pkts_polled))
1097 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1099 if (r_vec->nfp_net->rx_coalesce_adapt_on && r_vec->rx_ring) {
1100 struct dim_sample dim_sample = {};
1105 start = u64_stats_fetch_begin(&r_vec->rx_sync);
1106 pkts = r_vec->rx_pkts;
1107 bytes = r_vec->rx_bytes;
1108 } while (u64_stats_fetch_retry(&r_vec->rx_sync, start));
1110 dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1111 net_dim(&r_vec->rx_dim, dim_sample);
1114 if (r_vec->nfp_net->tx_coalesce_adapt_on && r_vec->tx_ring) {
1115 struct dim_sample dim_sample = {};
1120 start = u64_stats_fetch_begin(&r_vec->tx_sync);
1121 pkts = r_vec->tx_pkts;
1122 bytes = r_vec->tx_bytes;
1123 } while (u64_stats_fetch_retry(&r_vec->tx_sync, start));
1125 dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1126 net_dim(&r_vec->tx_dim, dim_sample);
1132 /* Control device data path
1136 nfp_nfd3_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1137 struct sk_buff *skb, bool old)
1139 unsigned int real_len = skb->len, meta_len = 0;
1140 struct nfp_net_tx_ring *tx_ring;
1141 struct nfp_nfd3_tx_buf *txbuf;
1142 struct nfp_nfd3_tx_desc *txd;
1143 struct nfp_net_dp *dp;
1144 dma_addr_t dma_addr;
1147 dp = &r_vec->nfp_net->dp;
1148 tx_ring = r_vec->tx_ring;
1150 if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1151 nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1155 if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1156 u64_stats_update_begin(&r_vec->tx_sync);
1158 u64_stats_update_end(&r_vec->tx_sync);
1160 __skb_queue_tail(&r_vec->queue, skb);
1162 __skb_queue_head(&r_vec->queue, skb);
1166 if (nfp_app_ctrl_has_meta(nn->app)) {
1167 if (unlikely(skb_headroom(skb) < 8)) {
1168 nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1172 put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1173 put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4));
1176 /* Start with the head skbuf */
1177 dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
1179 if (dma_mapping_error(dp->dev, dma_addr))
1182 wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1184 /* Stash the soft descriptor of the head then initialize it */
1185 txbuf = &tx_ring->txbufs[wr_idx];
1187 txbuf->dma_addr = dma_addr;
1190 txbuf->real_len = real_len;
1192 /* Build TX descriptor */
1193 txd = &tx_ring->txds[wr_idx];
1194 txd->offset_eop = meta_len | NFD3_DESC_TX_EOP;
1195 txd->dma_len = cpu_to_le16(skb_headlen(skb));
1196 nfp_desc_set_dma_addr(txd, dma_addr);
1197 txd->data_len = cpu_to_le16(skb->len);
1201 txd->lso_hdrlen = 0;
1204 tx_ring->wr_ptr_add++;
1205 nfp_net_tx_xmit_more_flush(tx_ring);
1210 nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n");
1212 u64_stats_update_begin(&r_vec->tx_sync);
1214 u64_stats_update_end(&r_vec->tx_sync);
1215 dev_kfree_skb_any(skb);
1219 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1221 struct sk_buff *skb;
1223 while ((skb = __skb_dequeue(&r_vec->queue)))
1224 if (nfp_nfd3_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1229 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1231 u32 meta_type, meta_tag;
1233 if (!nfp_app_ctrl_has_meta(nn->app))
1239 meta_type = get_unaligned_be32(data);
1240 meta_tag = get_unaligned_be32(data + 4);
1242 return (meta_type == NFP_NET_META_PORTID &&
1243 meta_tag == NFP_META_PORT_ID_CTRL);
1247 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1248 struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1250 unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1251 struct nfp_net_rx_buf *rxbuf;
1252 struct nfp_net_rx_desc *rxd;
1253 dma_addr_t new_dma_addr;
1254 struct sk_buff *skb;
1258 idx = D_IDX(rx_ring, rx_ring->rd_p);
1260 rxd = &rx_ring->rxds[idx];
1261 if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1264 /* Memory barrier to ensure that we won't do other reads
1265 * before the DD bit.
1271 rxbuf = &rx_ring->rxbufs[idx];
1272 meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1273 data_len = le16_to_cpu(rxd->rxd.data_len);
1274 pkt_len = data_len - meta_len;
1276 pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1277 if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1278 pkt_off += meta_len;
1280 pkt_off += dp->rx_offset;
1281 meta_off = pkt_off - meta_len;
1284 u64_stats_update_begin(&r_vec->rx_sync);
1286 r_vec->rx_bytes += pkt_len;
1287 u64_stats_update_end(&r_vec->rx_sync);
1289 nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len);
1291 if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
1292 nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
1294 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1298 skb = build_skb(rxbuf->frag, dp->fl_bufsz);
1299 if (unlikely(!skb)) {
1300 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1303 new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1304 if (unlikely(!new_frag)) {
1305 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1309 nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1311 nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1313 skb_reserve(skb, pkt_off);
1314 skb_put(skb, pkt_len);
1316 nfp_app_ctrl_rx(nn->app, skb);
1321 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
1323 struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
1324 struct nfp_net *nn = r_vec->nfp_net;
1325 struct nfp_net_dp *dp = &nn->dp;
1326 unsigned int budget = 512;
1328 while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
1334 void nfp_nfd3_ctrl_poll(struct tasklet_struct *t)
1336 struct nfp_net_r_vector *r_vec = from_tasklet(r_vec, t, tasklet);
1338 spin_lock(&r_vec->lock);
1339 nfp_nfd3_tx_complete(r_vec->tx_ring, 0);
1340 __nfp_ctrl_tx_queued(r_vec);
1341 spin_unlock(&r_vec->lock);
1343 if (nfp_ctrl_rx(r_vec)) {
1344 nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1346 tasklet_schedule(&r_vec->tasklet);
1347 nn_dp_warn(&r_vec->nfp_net->dp,
1348 "control message budget exceeded!\n");