Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause)
0002 /* Copyright (C) 2015-2019 Netronome Systems, Inc. */
0003 
0004 #include <linux/bpf_trace.h>
0005 #include <linux/netdevice.h>
0006 #include <linux/bitfield.h>
0007 
0008 #include "../nfp_app.h"
0009 #include "../nfp_net.h"
0010 #include "../nfp_net_dp.h"
0011 #include "../nfp_net_xsk.h"
0012 #include "../crypto/crypto.h"
0013 #include "../crypto/fw.h"
0014 #include "nfd3.h"
0015 
0016 /* Transmit processing
0017  *
0018  * One queue controller peripheral queue is used for transmit.  The
0019  * driver en-queues packets for transmit by advancing the write
0020  * pointer.  The device indicates that packets have transmitted by
0021  * advancing the read pointer.  The driver maintains a local copy of
0022  * the read and write pointer in @struct nfp_net_tx_ring.  The driver
0023  * keeps @wr_p in sync with the queue controller write pointer and can
0024  * determine how many packets have been transmitted by comparing its
0025  * copy of the read pointer @rd_p with the read pointer maintained by
0026  * the queue controller peripheral.
0027  */
0028 
0029 /* Wrappers for deciding when to stop and restart TX queues */
0030 static int nfp_nfd3_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
0031 {
0032     return !nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS * 4);
0033 }
0034 
0035 static int nfp_nfd3_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
0036 {
0037     return nfp_net_tx_full(tx_ring, MAX_SKB_FRAGS + 1);
0038 }
0039 
0040 /**
0041  * nfp_nfd3_tx_ring_stop() - stop tx ring
0042  * @nd_q:    netdev queue
0043  * @tx_ring: driver tx queue structure
0044  *
0045  * Safely stop TX ring.  Remember that while we are running .start_xmit()
0046  * someone else may be cleaning the TX ring completions so we need to be
0047  * extra careful here.
0048  */
0049 static void
0050 nfp_nfd3_tx_ring_stop(struct netdev_queue *nd_q,
0051               struct nfp_net_tx_ring *tx_ring)
0052 {
0053     netif_tx_stop_queue(nd_q);
0054 
0055     /* We can race with the TX completion out of NAPI so recheck */
0056     smp_mb();
0057     if (unlikely(nfp_nfd3_tx_ring_should_wake(tx_ring)))
0058         netif_tx_start_queue(nd_q);
0059 }
0060 
0061 /**
0062  * nfp_nfd3_tx_tso() - Set up Tx descriptor for LSO
0063  * @r_vec: per-ring structure
0064  * @txbuf: Pointer to driver soft TX descriptor
0065  * @txd: Pointer to HW TX descriptor
0066  * @skb: Pointer to SKB
0067  * @md_bytes: Prepend length
0068  *
0069  * Set up Tx descriptor for LSO, do nothing for non-LSO skbs.
0070  * Return error on packet header greater than maximum supported LSO header size.
0071  */
0072 static void
0073 nfp_nfd3_tx_tso(struct nfp_net_r_vector *r_vec, struct nfp_nfd3_tx_buf *txbuf,
0074         struct nfp_nfd3_tx_desc *txd, struct sk_buff *skb, u32 md_bytes)
0075 {
0076     u32 l3_offset, l4_offset, hdrlen;
0077     u16 mss;
0078 
0079     if (!skb_is_gso(skb))
0080         return;
0081 
0082     if (!skb->encapsulation) {
0083         l3_offset = skb_network_offset(skb);
0084         l4_offset = skb_transport_offset(skb);
0085         hdrlen = skb_tcp_all_headers(skb);
0086     } else {
0087         l3_offset = skb_inner_network_offset(skb);
0088         l4_offset = skb_inner_transport_offset(skb);
0089         hdrlen = skb_inner_tcp_all_headers(skb);
0090     }
0091 
0092     txbuf->pkt_cnt = skb_shinfo(skb)->gso_segs;
0093     txbuf->real_len += hdrlen * (txbuf->pkt_cnt - 1);
0094 
0095     mss = skb_shinfo(skb)->gso_size & NFD3_DESC_TX_MSS_MASK;
0096     txd->l3_offset = l3_offset - md_bytes;
0097     txd->l4_offset = l4_offset - md_bytes;
0098     txd->lso_hdrlen = hdrlen - md_bytes;
0099     txd->mss = cpu_to_le16(mss);
0100     txd->flags |= NFD3_DESC_TX_LSO;
0101 
0102     u64_stats_update_begin(&r_vec->tx_sync);
0103     r_vec->tx_lso++;
0104     u64_stats_update_end(&r_vec->tx_sync);
0105 }
0106 
0107 /**
0108  * nfp_nfd3_tx_csum() - Set TX CSUM offload flags in TX descriptor
0109  * @dp:  NFP Net data path struct
0110  * @r_vec: per-ring structure
0111  * @txbuf: Pointer to driver soft TX descriptor
0112  * @txd: Pointer to TX descriptor
0113  * @skb: Pointer to SKB
0114  *
0115  * This function sets the TX checksum flags in the TX descriptor based
0116  * on the configuration and the protocol of the packet to be transmitted.
0117  */
0118 static void
0119 nfp_nfd3_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
0120          struct nfp_nfd3_tx_buf *txbuf, struct nfp_nfd3_tx_desc *txd,
0121          struct sk_buff *skb)
0122 {
0123     struct ipv6hdr *ipv6h;
0124     struct iphdr *iph;
0125     u8 l4_hdr;
0126 
0127     if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
0128         return;
0129 
0130     if (skb->ip_summed != CHECKSUM_PARTIAL)
0131         return;
0132 
0133     txd->flags |= NFD3_DESC_TX_CSUM;
0134     if (skb->encapsulation)
0135         txd->flags |= NFD3_DESC_TX_ENCAP;
0136 
0137     iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
0138     ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
0139 
0140     if (iph->version == 4) {
0141         txd->flags |= NFD3_DESC_TX_IP4_CSUM;
0142         l4_hdr = iph->protocol;
0143     } else if (ipv6h->version == 6) {
0144         l4_hdr = ipv6h->nexthdr;
0145     } else {
0146         nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
0147         return;
0148     }
0149 
0150     switch (l4_hdr) {
0151     case IPPROTO_TCP:
0152         txd->flags |= NFD3_DESC_TX_TCP_CSUM;
0153         break;
0154     case IPPROTO_UDP:
0155         txd->flags |= NFD3_DESC_TX_UDP_CSUM;
0156         break;
0157     default:
0158         nn_dp_warn(dp, "partial checksum but l4 proto=%x!\n", l4_hdr);
0159         return;
0160     }
0161 
0162     u64_stats_update_begin(&r_vec->tx_sync);
0163     if (skb->encapsulation)
0164         r_vec->hw_csum_tx_inner += txbuf->pkt_cnt;
0165     else
0166         r_vec->hw_csum_tx += txbuf->pkt_cnt;
0167     u64_stats_update_end(&r_vec->tx_sync);
0168 }
0169 
0170 static int nfp_nfd3_prep_tx_meta(struct nfp_net_dp *dp, struct sk_buff *skb, u64 tls_handle)
0171 {
0172     struct metadata_dst *md_dst = skb_metadata_dst(skb);
0173     unsigned char *data;
0174     bool vlan_insert;
0175     u32 meta_id = 0;
0176     int md_bytes;
0177 
0178     if (unlikely(md_dst || tls_handle)) {
0179         if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX))
0180             md_dst = NULL;
0181     }
0182 
0183     vlan_insert = skb_vlan_tag_present(skb) && (dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN_V2);
0184 
0185     if (!(md_dst || tls_handle || vlan_insert))
0186         return 0;
0187 
0188     md_bytes = sizeof(meta_id) +
0189            !!md_dst * NFP_NET_META_PORTID_SIZE +
0190            !!tls_handle * NFP_NET_META_CONN_HANDLE_SIZE +
0191            vlan_insert * NFP_NET_META_VLAN_SIZE;
0192 
0193     if (unlikely(skb_cow_head(skb, md_bytes)))
0194         return -ENOMEM;
0195 
0196     data = skb_push(skb, md_bytes) + md_bytes;
0197     if (md_dst) {
0198         data -= NFP_NET_META_PORTID_SIZE;
0199         put_unaligned_be32(md_dst->u.port_info.port_id, data);
0200         meta_id = NFP_NET_META_PORTID;
0201     }
0202     if (tls_handle) {
0203         /* conn handle is opaque, we just use u64 to be able to quickly
0204          * compare it to zero
0205          */
0206         data -= NFP_NET_META_CONN_HANDLE_SIZE;
0207         memcpy(data, &tls_handle, sizeof(tls_handle));
0208         meta_id <<= NFP_NET_META_FIELD_SIZE;
0209         meta_id |= NFP_NET_META_CONN_HANDLE;
0210     }
0211     if (vlan_insert) {
0212         data -= NFP_NET_META_VLAN_SIZE;
0213         /* data type of skb->vlan_proto is __be16
0214          * so it fills metadata without calling put_unaligned_be16
0215          */
0216         memcpy(data, &skb->vlan_proto, sizeof(skb->vlan_proto));
0217         put_unaligned_be16(skb_vlan_tag_get(skb), data + sizeof(skb->vlan_proto));
0218         meta_id <<= NFP_NET_META_FIELD_SIZE;
0219         meta_id |= NFP_NET_META_VLAN;
0220     }
0221 
0222     data -= sizeof(meta_id);
0223     put_unaligned_be32(meta_id, data);
0224 
0225     return md_bytes;
0226 }
0227 
0228 /**
0229  * nfp_nfd3_tx() - Main transmit entry point
0230  * @skb:    SKB to transmit
0231  * @netdev: netdev structure
0232  *
0233  * Return: NETDEV_TX_OK on success.
0234  */
0235 netdev_tx_t nfp_nfd3_tx(struct sk_buff *skb, struct net_device *netdev)
0236 {
0237     struct nfp_net *nn = netdev_priv(netdev);
0238     int f, nr_frags, wr_idx, md_bytes;
0239     struct nfp_net_tx_ring *tx_ring;
0240     struct nfp_net_r_vector *r_vec;
0241     struct nfp_nfd3_tx_buf *txbuf;
0242     struct nfp_nfd3_tx_desc *txd;
0243     struct netdev_queue *nd_q;
0244     const skb_frag_t *frag;
0245     struct nfp_net_dp *dp;
0246     dma_addr_t dma_addr;
0247     unsigned int fsize;
0248     u64 tls_handle = 0;
0249     u16 qidx;
0250 
0251     dp = &nn->dp;
0252     qidx = skb_get_queue_mapping(skb);
0253     tx_ring = &dp->tx_rings[qidx];
0254     r_vec = tx_ring->r_vec;
0255 
0256     nr_frags = skb_shinfo(skb)->nr_frags;
0257 
0258     if (unlikely(nfp_net_tx_full(tx_ring, nr_frags + 1))) {
0259         nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
0260                qidx, tx_ring->wr_p, tx_ring->rd_p);
0261         nd_q = netdev_get_tx_queue(dp->netdev, qidx);
0262         netif_tx_stop_queue(nd_q);
0263         nfp_net_tx_xmit_more_flush(tx_ring);
0264         u64_stats_update_begin(&r_vec->tx_sync);
0265         r_vec->tx_busy++;
0266         u64_stats_update_end(&r_vec->tx_sync);
0267         return NETDEV_TX_BUSY;
0268     }
0269 
0270     skb = nfp_net_tls_tx(dp, r_vec, skb, &tls_handle, &nr_frags);
0271     if (unlikely(!skb)) {
0272         nfp_net_tx_xmit_more_flush(tx_ring);
0273         return NETDEV_TX_OK;
0274     }
0275 
0276     md_bytes = nfp_nfd3_prep_tx_meta(dp, skb, tls_handle);
0277     if (unlikely(md_bytes < 0))
0278         goto err_flush;
0279 
0280     /* Start with the head skbuf */
0281     dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
0282                   DMA_TO_DEVICE);
0283     if (dma_mapping_error(dp->dev, dma_addr))
0284         goto err_dma_err;
0285 
0286     wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
0287 
0288     /* Stash the soft descriptor of the head then initialize it */
0289     txbuf = &tx_ring->txbufs[wr_idx];
0290     txbuf->skb = skb;
0291     txbuf->dma_addr = dma_addr;
0292     txbuf->fidx = -1;
0293     txbuf->pkt_cnt = 1;
0294     txbuf->real_len = skb->len;
0295 
0296     /* Build TX descriptor */
0297     txd = &tx_ring->txds[wr_idx];
0298     txd->offset_eop = (nr_frags ? 0 : NFD3_DESC_TX_EOP) | md_bytes;
0299     txd->dma_len = cpu_to_le16(skb_headlen(skb));
0300     nfp_desc_set_dma_addr_40b(txd, dma_addr);
0301     txd->data_len = cpu_to_le16(skb->len);
0302 
0303     txd->flags = 0;
0304     txd->mss = 0;
0305     txd->lso_hdrlen = 0;
0306 
0307     /* Do not reorder - tso may adjust pkt cnt, vlan may override fields */
0308     nfp_nfd3_tx_tso(r_vec, txbuf, txd, skb, md_bytes);
0309     nfp_nfd3_tx_csum(dp, r_vec, txbuf, txd, skb);
0310     if (skb_vlan_tag_present(skb) && dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN) {
0311         txd->flags |= NFD3_DESC_TX_VLAN;
0312         txd->vlan = cpu_to_le16(skb_vlan_tag_get(skb));
0313     }
0314 
0315     /* Gather DMA */
0316     if (nr_frags > 0) {
0317         __le64 second_half;
0318 
0319         /* all descs must match except for in addr, length and eop */
0320         second_half = txd->vals8[1];
0321 
0322         for (f = 0; f < nr_frags; f++) {
0323             frag = &skb_shinfo(skb)->frags[f];
0324             fsize = skb_frag_size(frag);
0325 
0326             dma_addr = skb_frag_dma_map(dp->dev, frag, 0,
0327                             fsize, DMA_TO_DEVICE);
0328             if (dma_mapping_error(dp->dev, dma_addr))
0329                 goto err_unmap;
0330 
0331             wr_idx = D_IDX(tx_ring, wr_idx + 1);
0332             tx_ring->txbufs[wr_idx].skb = skb;
0333             tx_ring->txbufs[wr_idx].dma_addr = dma_addr;
0334             tx_ring->txbufs[wr_idx].fidx = f;
0335 
0336             txd = &tx_ring->txds[wr_idx];
0337             txd->dma_len = cpu_to_le16(fsize);
0338             nfp_desc_set_dma_addr_40b(txd, dma_addr);
0339             txd->offset_eop = md_bytes |
0340                 ((f == nr_frags - 1) ? NFD3_DESC_TX_EOP : 0);
0341             txd->vals8[1] = second_half;
0342         }
0343 
0344         u64_stats_update_begin(&r_vec->tx_sync);
0345         r_vec->tx_gather++;
0346         u64_stats_update_end(&r_vec->tx_sync);
0347     }
0348 
0349     skb_tx_timestamp(skb);
0350 
0351     nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
0352 
0353     tx_ring->wr_p += nr_frags + 1;
0354     if (nfp_nfd3_tx_ring_should_stop(tx_ring))
0355         nfp_nfd3_tx_ring_stop(nd_q, tx_ring);
0356 
0357     tx_ring->wr_ptr_add += nr_frags + 1;
0358     if (__netdev_tx_sent_queue(nd_q, txbuf->real_len, netdev_xmit_more()))
0359         nfp_net_tx_xmit_more_flush(tx_ring);
0360 
0361     return NETDEV_TX_OK;
0362 
0363 err_unmap:
0364     while (--f >= 0) {
0365         frag = &skb_shinfo(skb)->frags[f];
0366         dma_unmap_page(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
0367                    skb_frag_size(frag), DMA_TO_DEVICE);
0368         tx_ring->txbufs[wr_idx].skb = NULL;
0369         tx_ring->txbufs[wr_idx].dma_addr = 0;
0370         tx_ring->txbufs[wr_idx].fidx = -2;
0371         wr_idx = wr_idx - 1;
0372         if (wr_idx < 0)
0373             wr_idx += tx_ring->cnt;
0374     }
0375     dma_unmap_single(dp->dev, tx_ring->txbufs[wr_idx].dma_addr,
0376              skb_headlen(skb), DMA_TO_DEVICE);
0377     tx_ring->txbufs[wr_idx].skb = NULL;
0378     tx_ring->txbufs[wr_idx].dma_addr = 0;
0379     tx_ring->txbufs[wr_idx].fidx = -2;
0380 err_dma_err:
0381     nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
0382 err_flush:
0383     nfp_net_tx_xmit_more_flush(tx_ring);
0384     u64_stats_update_begin(&r_vec->tx_sync);
0385     r_vec->tx_errors++;
0386     u64_stats_update_end(&r_vec->tx_sync);
0387     nfp_net_tls_tx_undo(skb, tls_handle);
0388     dev_kfree_skb_any(skb);
0389     return NETDEV_TX_OK;
0390 }
0391 
0392 /**
0393  * nfp_nfd3_tx_complete() - Handled completed TX packets
0394  * @tx_ring:    TX ring structure
0395  * @budget: NAPI budget (only used as bool to determine if in NAPI context)
0396  */
0397 void nfp_nfd3_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
0398 {
0399     struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
0400     struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
0401     u32 done_pkts = 0, done_bytes = 0;
0402     struct netdev_queue *nd_q;
0403     u32 qcp_rd_p;
0404     int todo;
0405 
0406     if (tx_ring->wr_p == tx_ring->rd_p)
0407         return;
0408 
0409     /* Work out how many descriptors have been transmitted */
0410     qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
0411 
0412     if (qcp_rd_p == tx_ring->qcp_rd_p)
0413         return;
0414 
0415     todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
0416 
0417     while (todo--) {
0418         const skb_frag_t *frag;
0419         struct nfp_nfd3_tx_buf *tx_buf;
0420         struct sk_buff *skb;
0421         int fidx, nr_frags;
0422         int idx;
0423 
0424         idx = D_IDX(tx_ring, tx_ring->rd_p++);
0425         tx_buf = &tx_ring->txbufs[idx];
0426 
0427         skb = tx_buf->skb;
0428         if (!skb)
0429             continue;
0430 
0431         nr_frags = skb_shinfo(skb)->nr_frags;
0432         fidx = tx_buf->fidx;
0433 
0434         if (fidx == -1) {
0435             /* unmap head */
0436             dma_unmap_single(dp->dev, tx_buf->dma_addr,
0437                      skb_headlen(skb), DMA_TO_DEVICE);
0438 
0439             done_pkts += tx_buf->pkt_cnt;
0440             done_bytes += tx_buf->real_len;
0441         } else {
0442             /* unmap fragment */
0443             frag = &skb_shinfo(skb)->frags[fidx];
0444             dma_unmap_page(dp->dev, tx_buf->dma_addr,
0445                        skb_frag_size(frag), DMA_TO_DEVICE);
0446         }
0447 
0448         /* check for last gather fragment */
0449         if (fidx == nr_frags - 1)
0450             napi_consume_skb(skb, budget);
0451 
0452         tx_buf->dma_addr = 0;
0453         tx_buf->skb = NULL;
0454         tx_buf->fidx = -2;
0455     }
0456 
0457     tx_ring->qcp_rd_p = qcp_rd_p;
0458 
0459     u64_stats_update_begin(&r_vec->tx_sync);
0460     r_vec->tx_bytes += done_bytes;
0461     r_vec->tx_pkts += done_pkts;
0462     u64_stats_update_end(&r_vec->tx_sync);
0463 
0464     if (!dp->netdev)
0465         return;
0466 
0467     nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
0468     netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
0469     if (nfp_nfd3_tx_ring_should_wake(tx_ring)) {
0470         /* Make sure TX thread will see updated tx_ring->rd_p */
0471         smp_mb();
0472 
0473         if (unlikely(netif_tx_queue_stopped(nd_q)))
0474             netif_tx_wake_queue(nd_q);
0475     }
0476 
0477     WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
0478           "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
0479           tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
0480 }
0481 
0482 static bool nfp_nfd3_xdp_complete(struct nfp_net_tx_ring *tx_ring)
0483 {
0484     struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
0485     struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
0486     u32 done_pkts = 0, done_bytes = 0;
0487     bool done_all;
0488     int idx, todo;
0489     u32 qcp_rd_p;
0490 
0491     /* Work out how many descriptors have been transmitted */
0492     qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
0493 
0494     if (qcp_rd_p == tx_ring->qcp_rd_p)
0495         return true;
0496 
0497     todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
0498 
0499     done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
0500     todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
0501 
0502     tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + todo);
0503 
0504     done_pkts = todo;
0505     while (todo--) {
0506         idx = D_IDX(tx_ring, tx_ring->rd_p);
0507         tx_ring->rd_p++;
0508 
0509         done_bytes += tx_ring->txbufs[idx].real_len;
0510     }
0511 
0512     u64_stats_update_begin(&r_vec->tx_sync);
0513     r_vec->tx_bytes += done_bytes;
0514     r_vec->tx_pkts += done_pkts;
0515     u64_stats_update_end(&r_vec->tx_sync);
0516 
0517     WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
0518           "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
0519           tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
0520 
0521     return done_all;
0522 }
0523 
0524 /* Receive processing
0525  */
0526 
0527 static void *
0528 nfp_nfd3_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
0529 {
0530     void *frag;
0531 
0532     if (!dp->xdp_prog) {
0533         frag = napi_alloc_frag(dp->fl_bufsz);
0534         if (unlikely(!frag))
0535             return NULL;
0536     } else {
0537         struct page *page;
0538 
0539         page = dev_alloc_page();
0540         if (unlikely(!page))
0541             return NULL;
0542         frag = page_address(page);
0543     }
0544 
0545     *dma_addr = nfp_net_dma_map_rx(dp, frag);
0546     if (dma_mapping_error(dp->dev, *dma_addr)) {
0547         nfp_net_free_frag(frag, dp->xdp_prog);
0548         nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
0549         return NULL;
0550     }
0551 
0552     return frag;
0553 }
0554 
0555 /**
0556  * nfp_nfd3_rx_give_one() - Put mapped skb on the software and hardware rings
0557  * @dp:     NFP Net data path struct
0558  * @rx_ring:    RX ring structure
0559  * @frag:   page fragment buffer
0560  * @dma_addr:   DMA address of skb mapping
0561  */
0562 static void
0563 nfp_nfd3_rx_give_one(const struct nfp_net_dp *dp,
0564              struct nfp_net_rx_ring *rx_ring,
0565              void *frag, dma_addr_t dma_addr)
0566 {
0567     unsigned int wr_idx;
0568 
0569     wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
0570 
0571     nfp_net_dma_sync_dev_rx(dp, dma_addr);
0572 
0573     /* Stash SKB and DMA address away */
0574     rx_ring->rxbufs[wr_idx].frag = frag;
0575     rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
0576 
0577     /* Fill freelist descriptor */
0578     rx_ring->rxds[wr_idx].fld.reserved = 0;
0579     rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
0580     /* DMA address is expanded to 48-bit width in freelist for NFP3800,
0581      * so the *_48b macro is used accordingly, it's also OK to fill
0582      * a 40-bit address since the top 8 bits are get set to 0.
0583      */
0584     nfp_desc_set_dma_addr_48b(&rx_ring->rxds[wr_idx].fld,
0585                   dma_addr + dp->rx_dma_off);
0586 
0587     rx_ring->wr_p++;
0588     if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
0589         /* Update write pointer of the freelist queue. Make
0590          * sure all writes are flushed before telling the hardware.
0591          */
0592         wmb();
0593         nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
0594     }
0595 }
0596 
0597 /**
0598  * nfp_nfd3_rx_ring_fill_freelist() - Give buffers from the ring to FW
0599  * @dp:      NFP Net data path struct
0600  * @rx_ring: RX ring to fill
0601  */
0602 void nfp_nfd3_rx_ring_fill_freelist(struct nfp_net_dp *dp,
0603                     struct nfp_net_rx_ring *rx_ring)
0604 {
0605     unsigned int i;
0606 
0607     if (nfp_net_has_xsk_pool_slow(dp, rx_ring->idx))
0608         return nfp_net_xsk_rx_ring_fill_freelist(rx_ring);
0609 
0610     for (i = 0; i < rx_ring->cnt - 1; i++)
0611         nfp_nfd3_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
0612                      rx_ring->rxbufs[i].dma_addr);
0613 }
0614 
0615 /**
0616  * nfp_nfd3_rx_csum_has_errors() - group check if rxd has any csum errors
0617  * @flags: RX descriptor flags field in CPU byte order
0618  */
0619 static int nfp_nfd3_rx_csum_has_errors(u16 flags)
0620 {
0621     u16 csum_all_checked, csum_all_ok;
0622 
0623     csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
0624     csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
0625 
0626     return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
0627 }
0628 
0629 /**
0630  * nfp_nfd3_rx_csum() - set SKB checksum field based on RX descriptor flags
0631  * @dp:  NFP Net data path struct
0632  * @r_vec: per-ring structure
0633  * @rxd: Pointer to RX descriptor
0634  * @meta: Parsed metadata prepend
0635  * @skb: Pointer to SKB
0636  */
0637 void
0638 nfp_nfd3_rx_csum(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
0639          const struct nfp_net_rx_desc *rxd,
0640          const struct nfp_meta_parsed *meta, struct sk_buff *skb)
0641 {
0642     skb_checksum_none_assert(skb);
0643 
0644     if (!(dp->netdev->features & NETIF_F_RXCSUM))
0645         return;
0646 
0647     if (meta->csum_type) {
0648         skb->ip_summed = meta->csum_type;
0649         skb->csum = meta->csum;
0650         u64_stats_update_begin(&r_vec->rx_sync);
0651         r_vec->hw_csum_rx_complete++;
0652         u64_stats_update_end(&r_vec->rx_sync);
0653         return;
0654     }
0655 
0656     if (nfp_nfd3_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
0657         u64_stats_update_begin(&r_vec->rx_sync);
0658         r_vec->hw_csum_rx_error++;
0659         u64_stats_update_end(&r_vec->rx_sync);
0660         return;
0661     }
0662 
0663     /* Assume that the firmware will never report inner CSUM_OK unless outer
0664      * L4 headers were successfully parsed. FW will always report zero UDP
0665      * checksum as CSUM_OK.
0666      */
0667     if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
0668         rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
0669         __skb_incr_checksum_unnecessary(skb);
0670         u64_stats_update_begin(&r_vec->rx_sync);
0671         r_vec->hw_csum_rx_ok++;
0672         u64_stats_update_end(&r_vec->rx_sync);
0673     }
0674 
0675     if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
0676         rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
0677         __skb_incr_checksum_unnecessary(skb);
0678         u64_stats_update_begin(&r_vec->rx_sync);
0679         r_vec->hw_csum_rx_inner_ok++;
0680         u64_stats_update_end(&r_vec->rx_sync);
0681     }
0682 }
0683 
0684 static void
0685 nfp_nfd3_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
0686           unsigned int type, __be32 *hash)
0687 {
0688     if (!(netdev->features & NETIF_F_RXHASH))
0689         return;
0690 
0691     switch (type) {
0692     case NFP_NET_RSS_IPV4:
0693     case NFP_NET_RSS_IPV6:
0694     case NFP_NET_RSS_IPV6_EX:
0695         meta->hash_type = PKT_HASH_TYPE_L3;
0696         break;
0697     default:
0698         meta->hash_type = PKT_HASH_TYPE_L4;
0699         break;
0700     }
0701 
0702     meta->hash = get_unaligned_be32(hash);
0703 }
0704 
0705 static void
0706 nfp_nfd3_set_hash_desc(struct net_device *netdev, struct nfp_meta_parsed *meta,
0707                void *data, struct nfp_net_rx_desc *rxd)
0708 {
0709     struct nfp_net_rx_hash *rx_hash = data;
0710 
0711     if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
0712         return;
0713 
0714     nfp_nfd3_set_hash(netdev, meta, get_unaligned_be32(&rx_hash->hash_type),
0715               &rx_hash->hash);
0716 }
0717 
0718 bool
0719 nfp_nfd3_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
0720             void *data, void *pkt, unsigned int pkt_len, int meta_len)
0721 {
0722     u32 meta_info, vlan_info;
0723 
0724     meta_info = get_unaligned_be32(data);
0725     data += 4;
0726 
0727     while (meta_info) {
0728         switch (meta_info & NFP_NET_META_FIELD_MASK) {
0729         case NFP_NET_META_HASH:
0730             meta_info >>= NFP_NET_META_FIELD_SIZE;
0731             nfp_nfd3_set_hash(netdev, meta,
0732                       meta_info & NFP_NET_META_FIELD_MASK,
0733                       (__be32 *)data);
0734             data += 4;
0735             break;
0736         case NFP_NET_META_MARK:
0737             meta->mark = get_unaligned_be32(data);
0738             data += 4;
0739             break;
0740         case NFP_NET_META_VLAN:
0741             vlan_info = get_unaligned_be32(data);
0742             if (FIELD_GET(NFP_NET_META_VLAN_STRIP, vlan_info)) {
0743                 meta->vlan.stripped = true;
0744                 meta->vlan.tpid = FIELD_GET(NFP_NET_META_VLAN_TPID_MASK,
0745                                 vlan_info);
0746                 meta->vlan.tci = FIELD_GET(NFP_NET_META_VLAN_TCI_MASK,
0747                                vlan_info);
0748             }
0749             data += 4;
0750             break;
0751         case NFP_NET_META_PORTID:
0752             meta->portid = get_unaligned_be32(data);
0753             data += 4;
0754             break;
0755         case NFP_NET_META_CSUM:
0756             meta->csum_type = CHECKSUM_COMPLETE;
0757             meta->csum =
0758                 (__force __wsum)__get_unaligned_cpu32(data);
0759             data += 4;
0760             break;
0761         case NFP_NET_META_RESYNC_INFO:
0762             if (nfp_net_tls_rx_resync_req(netdev, data, pkt,
0763                               pkt_len))
0764                 return false;
0765             data += sizeof(struct nfp_net_tls_resync_req);
0766             break;
0767         default:
0768             return true;
0769         }
0770 
0771         meta_info >>= NFP_NET_META_FIELD_SIZE;
0772     }
0773 
0774     return data != pkt;
0775 }
0776 
0777 static void
0778 nfp_nfd3_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
0779          struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf,
0780          struct sk_buff *skb)
0781 {
0782     u64_stats_update_begin(&r_vec->rx_sync);
0783     r_vec->rx_drops++;
0784     /* If we have both skb and rxbuf the replacement buffer allocation
0785      * must have failed, count this as an alloc failure.
0786      */
0787     if (skb && rxbuf)
0788         r_vec->rx_replace_buf_alloc_fail++;
0789     u64_stats_update_end(&r_vec->rx_sync);
0790 
0791     /* skb is build based on the frag, free_skb() would free the frag
0792      * so to be able to reuse it we need an extra ref.
0793      */
0794     if (skb && rxbuf && skb->head == rxbuf->frag)
0795         page_ref_inc(virt_to_head_page(rxbuf->frag));
0796     if (rxbuf)
0797         nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
0798     if (skb)
0799         dev_kfree_skb_any(skb);
0800 }
0801 
0802 static bool
0803 nfp_nfd3_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
0804             struct nfp_net_tx_ring *tx_ring,
0805             struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
0806             unsigned int pkt_len, bool *completed)
0807 {
0808     unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA;
0809     struct nfp_nfd3_tx_buf *txbuf;
0810     struct nfp_nfd3_tx_desc *txd;
0811     int wr_idx;
0812 
0813     /* Reject if xdp_adjust_tail grow packet beyond DMA area */
0814     if (pkt_len + dma_off > dma_map_sz)
0815         return false;
0816 
0817     if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
0818         if (!*completed) {
0819             nfp_nfd3_xdp_complete(tx_ring);
0820             *completed = true;
0821         }
0822 
0823         if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
0824             nfp_nfd3_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
0825                      NULL);
0826             return false;
0827         }
0828     }
0829 
0830     wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
0831 
0832     /* Stash the soft descriptor of the head then initialize it */
0833     txbuf = &tx_ring->txbufs[wr_idx];
0834 
0835     nfp_nfd3_rx_give_one(dp, rx_ring, txbuf->frag, txbuf->dma_addr);
0836 
0837     txbuf->frag = rxbuf->frag;
0838     txbuf->dma_addr = rxbuf->dma_addr;
0839     txbuf->fidx = -1;
0840     txbuf->pkt_cnt = 1;
0841     txbuf->real_len = pkt_len;
0842 
0843     dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
0844                    pkt_len, DMA_BIDIRECTIONAL);
0845 
0846     /* Build TX descriptor */
0847     txd = &tx_ring->txds[wr_idx];
0848     txd->offset_eop = NFD3_DESC_TX_EOP;
0849     txd->dma_len = cpu_to_le16(pkt_len);
0850     nfp_desc_set_dma_addr_40b(txd, rxbuf->dma_addr + dma_off);
0851     txd->data_len = cpu_to_le16(pkt_len);
0852 
0853     txd->flags = 0;
0854     txd->mss = 0;
0855     txd->lso_hdrlen = 0;
0856 
0857     tx_ring->wr_p++;
0858     tx_ring->wr_ptr_add++;
0859     return true;
0860 }
0861 
0862 /**
0863  * nfp_nfd3_rx() - receive up to @budget packets on @rx_ring
0864  * @rx_ring:   RX ring to receive from
0865  * @budget:    NAPI budget
0866  *
0867  * Note, this function is separated out from the napi poll function to
0868  * more cleanly separate packet receive code from other bookkeeping
0869  * functions performed in the napi poll function.
0870  *
0871  * Return: Number of packets received.
0872  */
0873 static int nfp_nfd3_rx(struct nfp_net_rx_ring *rx_ring, int budget)
0874 {
0875     struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
0876     struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
0877     struct nfp_net_tx_ring *tx_ring;
0878     struct bpf_prog *xdp_prog;
0879     bool xdp_tx_cmpl = false;
0880     unsigned int true_bufsz;
0881     struct sk_buff *skb;
0882     int pkts_polled = 0;
0883     struct xdp_buff xdp;
0884     int idx;
0885 
0886     xdp_prog = READ_ONCE(dp->xdp_prog);
0887     true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
0888     xdp_init_buff(&xdp, PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM,
0889               &rx_ring->xdp_rxq);
0890     tx_ring = r_vec->xdp_ring;
0891 
0892     while (pkts_polled < budget) {
0893         unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
0894         struct nfp_net_rx_buf *rxbuf;
0895         struct nfp_net_rx_desc *rxd;
0896         struct nfp_meta_parsed meta;
0897         bool redir_egress = false;
0898         struct net_device *netdev;
0899         dma_addr_t new_dma_addr;
0900         u32 meta_len_xdp = 0;
0901         void *new_frag;
0902 
0903         idx = D_IDX(rx_ring, rx_ring->rd_p);
0904 
0905         rxd = &rx_ring->rxds[idx];
0906         if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
0907             break;
0908 
0909         /* Memory barrier to ensure that we won't do other reads
0910          * before the DD bit.
0911          */
0912         dma_rmb();
0913 
0914         memset(&meta, 0, sizeof(meta));
0915 
0916         rx_ring->rd_p++;
0917         pkts_polled++;
0918 
0919         rxbuf = &rx_ring->rxbufs[idx];
0920         /*         < meta_len >
0921          *  <-- [rx_offset] -->
0922          *  ---------------------------------------------------------
0923          * | [XX] |  metadata  |             packet           | XXXX |
0924          *  ---------------------------------------------------------
0925          *         <---------------- data_len --------------->
0926          *
0927          * The rx_offset is fixed for all packets, the meta_len can vary
0928          * on a packet by packet basis. If rx_offset is set to zero
0929          * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
0930          * buffer and is immediately followed by the packet (no [XX]).
0931          */
0932         meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
0933         data_len = le16_to_cpu(rxd->rxd.data_len);
0934         pkt_len = data_len - meta_len;
0935 
0936         pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
0937         if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
0938             pkt_off += meta_len;
0939         else
0940             pkt_off += dp->rx_offset;
0941         meta_off = pkt_off - meta_len;
0942 
0943         /* Stats update */
0944         u64_stats_update_begin(&r_vec->rx_sync);
0945         r_vec->rx_pkts++;
0946         r_vec->rx_bytes += pkt_len;
0947         u64_stats_update_end(&r_vec->rx_sync);
0948 
0949         if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
0950                  (dp->rx_offset && meta_len > dp->rx_offset))) {
0951             nn_dp_warn(dp, "oversized RX packet metadata %u\n",
0952                    meta_len);
0953             nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
0954             continue;
0955         }
0956 
0957         nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
0958                     data_len);
0959 
0960         if (!dp->chained_metadata_format) {
0961             nfp_nfd3_set_hash_desc(dp->netdev, &meta,
0962                            rxbuf->frag + meta_off, rxd);
0963         } else if (meta_len) {
0964             if (unlikely(nfp_nfd3_parse_meta(dp->netdev, &meta,
0965                              rxbuf->frag + meta_off,
0966                              rxbuf->frag + pkt_off,
0967                              pkt_len, meta_len))) {
0968                 nn_dp_warn(dp, "invalid RX packet metadata\n");
0969                 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
0970                          NULL);
0971                 continue;
0972             }
0973         }
0974 
0975         if (xdp_prog && !meta.portid) {
0976             void *orig_data = rxbuf->frag + pkt_off;
0977             unsigned int dma_off;
0978             int act;
0979 
0980             xdp_prepare_buff(&xdp,
0981                      rxbuf->frag + NFP_NET_RX_BUF_HEADROOM,
0982                      pkt_off - NFP_NET_RX_BUF_HEADROOM,
0983                      pkt_len, true);
0984 
0985             act = bpf_prog_run_xdp(xdp_prog, &xdp);
0986 
0987             pkt_len = xdp.data_end - xdp.data;
0988             pkt_off += xdp.data - orig_data;
0989 
0990             switch (act) {
0991             case XDP_PASS:
0992                 meta_len_xdp = xdp.data - xdp.data_meta;
0993                 break;
0994             case XDP_TX:
0995                 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
0996                 if (unlikely(!nfp_nfd3_tx_xdp_buf(dp, rx_ring,
0997                                   tx_ring,
0998                                   rxbuf,
0999                                   dma_off,
1000                                   pkt_len,
1001                                   &xdp_tx_cmpl)))
1002                     trace_xdp_exception(dp->netdev,
1003                                 xdp_prog, act);
1004                 continue;
1005             default:
1006                 bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
1007                 fallthrough;
1008             case XDP_ABORTED:
1009                 trace_xdp_exception(dp->netdev, xdp_prog, act);
1010                 fallthrough;
1011             case XDP_DROP:
1012                 nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
1013                              rxbuf->dma_addr);
1014                 continue;
1015             }
1016         }
1017 
1018         if (likely(!meta.portid)) {
1019             netdev = dp->netdev;
1020         } else if (meta.portid == NFP_META_PORT_ID_CTRL) {
1021             struct nfp_net *nn = netdev_priv(dp->netdev);
1022 
1023             nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
1024                         pkt_len);
1025             nfp_nfd3_rx_give_one(dp, rx_ring, rxbuf->frag,
1026                          rxbuf->dma_addr);
1027             continue;
1028         } else {
1029             struct nfp_net *nn;
1030 
1031             nn = netdev_priv(dp->netdev);
1032             netdev = nfp_app_dev_get(nn->app, meta.portid,
1033                          &redir_egress);
1034             if (unlikely(!netdev)) {
1035                 nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf,
1036                          NULL);
1037                 continue;
1038             }
1039 
1040             if (nfp_netdev_is_nfp_repr(netdev))
1041                 nfp_repr_inc_rx_stats(netdev, pkt_len);
1042         }
1043 
1044         skb = build_skb(rxbuf->frag, true_bufsz);
1045         if (unlikely(!skb)) {
1046             nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1047             continue;
1048         }
1049         new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1050         if (unlikely(!new_frag)) {
1051             nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1052             continue;
1053         }
1054 
1055         nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1056 
1057         nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1058 
1059         skb_reserve(skb, pkt_off);
1060         skb_put(skb, pkt_len);
1061 
1062         skb->mark = meta.mark;
1063         skb_set_hash(skb, meta.hash, meta.hash_type);
1064 
1065         skb_record_rx_queue(skb, rx_ring->idx);
1066         skb->protocol = eth_type_trans(skb, netdev);
1067 
1068         nfp_nfd3_rx_csum(dp, r_vec, rxd, &meta, skb);
1069 
1070 #ifdef CONFIG_TLS_DEVICE
1071         if (rxd->rxd.flags & PCIE_DESC_RX_DECRYPTED) {
1072             skb->decrypted = true;
1073             u64_stats_update_begin(&r_vec->rx_sync);
1074             r_vec->hw_tls_rx++;
1075             u64_stats_update_end(&r_vec->rx_sync);
1076         }
1077 #endif
1078 
1079         if (unlikely(!nfp_net_vlan_strip(skb, rxd, &meta))) {
1080             nfp_nfd3_rx_drop(dp, r_vec, rx_ring, NULL, skb);
1081             continue;
1082         }
1083 
1084         if (meta_len_xdp)
1085             skb_metadata_set(skb, meta_len_xdp);
1086 
1087         if (likely(!redir_egress)) {
1088             napi_gro_receive(&rx_ring->r_vec->napi, skb);
1089         } else {
1090             skb->dev = netdev;
1091             skb_reset_network_header(skb);
1092             __skb_push(skb, ETH_HLEN);
1093             dev_queue_xmit(skb);
1094         }
1095     }
1096 
1097     if (xdp_prog) {
1098         if (tx_ring->wr_ptr_add)
1099             nfp_net_tx_xmit_more_flush(tx_ring);
1100         else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1101              !xdp_tx_cmpl)
1102             if (!nfp_nfd3_xdp_complete(tx_ring))
1103                 pkts_polled = budget;
1104     }
1105 
1106     return pkts_polled;
1107 }
1108 
1109 /**
1110  * nfp_nfd3_poll() - napi poll function
1111  * @napi:    NAPI structure
1112  * @budget:  NAPI budget
1113  *
1114  * Return: number of packets polled.
1115  */
1116 int nfp_nfd3_poll(struct napi_struct *napi, int budget)
1117 {
1118     struct nfp_net_r_vector *r_vec =
1119         container_of(napi, struct nfp_net_r_vector, napi);
1120     unsigned int pkts_polled = 0;
1121 
1122     if (r_vec->tx_ring)
1123         nfp_nfd3_tx_complete(r_vec->tx_ring, budget);
1124     if (r_vec->rx_ring)
1125         pkts_polled = nfp_nfd3_rx(r_vec->rx_ring, budget);
1126 
1127     if (pkts_polled < budget)
1128         if (napi_complete_done(napi, pkts_polled))
1129             nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1130 
1131     if (r_vec->nfp_net->rx_coalesce_adapt_on && r_vec->rx_ring) {
1132         struct dim_sample dim_sample = {};
1133         unsigned int start;
1134         u64 pkts, bytes;
1135 
1136         do {
1137             start = u64_stats_fetch_begin(&r_vec->rx_sync);
1138             pkts = r_vec->rx_pkts;
1139             bytes = r_vec->rx_bytes;
1140         } while (u64_stats_fetch_retry(&r_vec->rx_sync, start));
1141 
1142         dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1143         net_dim(&r_vec->rx_dim, dim_sample);
1144     }
1145 
1146     if (r_vec->nfp_net->tx_coalesce_adapt_on && r_vec->tx_ring) {
1147         struct dim_sample dim_sample = {};
1148         unsigned int start;
1149         u64 pkts, bytes;
1150 
1151         do {
1152             start = u64_stats_fetch_begin(&r_vec->tx_sync);
1153             pkts = r_vec->tx_pkts;
1154             bytes = r_vec->tx_bytes;
1155         } while (u64_stats_fetch_retry(&r_vec->tx_sync, start));
1156 
1157         dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1158         net_dim(&r_vec->tx_dim, dim_sample);
1159     }
1160 
1161     return pkts_polled;
1162 }
1163 
1164 /* Control device data path
1165  */
1166 
1167 bool
1168 nfp_nfd3_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1169              struct sk_buff *skb, bool old)
1170 {
1171     unsigned int real_len = skb->len, meta_len = 0;
1172     struct nfp_net_tx_ring *tx_ring;
1173     struct nfp_nfd3_tx_buf *txbuf;
1174     struct nfp_nfd3_tx_desc *txd;
1175     struct nfp_net_dp *dp;
1176     dma_addr_t dma_addr;
1177     int wr_idx;
1178 
1179     dp = &r_vec->nfp_net->dp;
1180     tx_ring = r_vec->tx_ring;
1181 
1182     if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1183         nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1184         goto err_free;
1185     }
1186 
1187     if (unlikely(nfp_net_tx_full(tx_ring, 1))) {
1188         u64_stats_update_begin(&r_vec->tx_sync);
1189         r_vec->tx_busy++;
1190         u64_stats_update_end(&r_vec->tx_sync);
1191         if (!old)
1192             __skb_queue_tail(&r_vec->queue, skb);
1193         else
1194             __skb_queue_head(&r_vec->queue, skb);
1195         return true;
1196     }
1197 
1198     if (nfp_app_ctrl_has_meta(nn->app)) {
1199         if (unlikely(skb_headroom(skb) < 8)) {
1200             nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1201             goto err_free;
1202         }
1203         meta_len = 8;
1204         put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1205         put_unaligned_be32(NFP_NET_META_PORTID, skb_push(skb, 4));
1206     }
1207 
1208     /* Start with the head skbuf */
1209     dma_addr = dma_map_single(dp->dev, skb->data, skb_headlen(skb),
1210                   DMA_TO_DEVICE);
1211     if (dma_mapping_error(dp->dev, dma_addr))
1212         goto err_dma_warn;
1213 
1214     wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1215 
1216     /* Stash the soft descriptor of the head then initialize it */
1217     txbuf = &tx_ring->txbufs[wr_idx];
1218     txbuf->skb = skb;
1219     txbuf->dma_addr = dma_addr;
1220     txbuf->fidx = -1;
1221     txbuf->pkt_cnt = 1;
1222     txbuf->real_len = real_len;
1223 
1224     /* Build TX descriptor */
1225     txd = &tx_ring->txds[wr_idx];
1226     txd->offset_eop = meta_len | NFD3_DESC_TX_EOP;
1227     txd->dma_len = cpu_to_le16(skb_headlen(skb));
1228     nfp_desc_set_dma_addr_40b(txd, dma_addr);
1229     txd->data_len = cpu_to_le16(skb->len);
1230 
1231     txd->flags = 0;
1232     txd->mss = 0;
1233     txd->lso_hdrlen = 0;
1234 
1235     tx_ring->wr_p++;
1236     tx_ring->wr_ptr_add++;
1237     nfp_net_tx_xmit_more_flush(tx_ring);
1238 
1239     return false;
1240 
1241 err_dma_warn:
1242     nn_dp_warn(dp, "Failed to DMA map TX CTRL buffer\n");
1243 err_free:
1244     u64_stats_update_begin(&r_vec->tx_sync);
1245     r_vec->tx_errors++;
1246     u64_stats_update_end(&r_vec->tx_sync);
1247     dev_kfree_skb_any(skb);
1248     return false;
1249 }
1250 
1251 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1252 {
1253     struct sk_buff *skb;
1254 
1255     while ((skb = __skb_dequeue(&r_vec->queue)))
1256         if (nfp_nfd3_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1257             return;
1258 }
1259 
1260 static bool
1261 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1262 {
1263     u32 meta_type, meta_tag;
1264 
1265     if (!nfp_app_ctrl_has_meta(nn->app))
1266         return !meta_len;
1267 
1268     if (meta_len != 8)
1269         return false;
1270 
1271     meta_type = get_unaligned_be32(data);
1272     meta_tag = get_unaligned_be32(data + 4);
1273 
1274     return (meta_type == NFP_NET_META_PORTID &&
1275         meta_tag == NFP_META_PORT_ID_CTRL);
1276 }
1277 
1278 static bool
1279 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1280         struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1281 {
1282     unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1283     struct nfp_net_rx_buf *rxbuf;
1284     struct nfp_net_rx_desc *rxd;
1285     dma_addr_t new_dma_addr;
1286     struct sk_buff *skb;
1287     void *new_frag;
1288     int idx;
1289 
1290     idx = D_IDX(rx_ring, rx_ring->rd_p);
1291 
1292     rxd = &rx_ring->rxds[idx];
1293     if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1294         return false;
1295 
1296     /* Memory barrier to ensure that we won't do other reads
1297      * before the DD bit.
1298      */
1299     dma_rmb();
1300 
1301     rx_ring->rd_p++;
1302 
1303     rxbuf = &rx_ring->rxbufs[idx];
1304     meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1305     data_len = le16_to_cpu(rxd->rxd.data_len);
1306     pkt_len = data_len - meta_len;
1307 
1308     pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1309     if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1310         pkt_off += meta_len;
1311     else
1312         pkt_off += dp->rx_offset;
1313     meta_off = pkt_off - meta_len;
1314 
1315     /* Stats update */
1316     u64_stats_update_begin(&r_vec->rx_sync);
1317     r_vec->rx_pkts++;
1318     r_vec->rx_bytes += pkt_len;
1319     u64_stats_update_end(&r_vec->rx_sync);
1320 
1321     nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len);
1322 
1323     if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
1324         nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
1325                meta_len);
1326         nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1327         return true;
1328     }
1329 
1330     skb = build_skb(rxbuf->frag, dp->fl_bufsz);
1331     if (unlikely(!skb)) {
1332         nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1333         return true;
1334     }
1335     new_frag = nfp_nfd3_napi_alloc_one(dp, &new_dma_addr);
1336     if (unlikely(!new_frag)) {
1337         nfp_nfd3_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1338         return true;
1339     }
1340 
1341     nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1342 
1343     nfp_nfd3_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1344 
1345     skb_reserve(skb, pkt_off);
1346     skb_put(skb, pkt_len);
1347 
1348     nfp_app_ctrl_rx(nn->app, skb);
1349 
1350     return true;
1351 }
1352 
1353 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
1354 {
1355     struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
1356     struct nfp_net *nn = r_vec->nfp_net;
1357     struct nfp_net_dp *dp = &nn->dp;
1358     unsigned int budget = 512;
1359 
1360     while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
1361         continue;
1362 
1363     return budget;
1364 }
1365 
1366 void nfp_nfd3_ctrl_poll(struct tasklet_struct *t)
1367 {
1368     struct nfp_net_r_vector *r_vec = from_tasklet(r_vec, t, tasklet);
1369 
1370     spin_lock(&r_vec->lock);
1371     nfp_nfd3_tx_complete(r_vec->tx_ring, 0);
1372     __nfp_ctrl_tx_queued(r_vec);
1373     spin_unlock(&r_vec->lock);
1374 
1375     if (nfp_ctrl_rx(r_vec)) {
1376         nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1377     } else {
1378         tasklet_schedule(&r_vec->tasklet);
1379         nn_dp_warn(&r_vec->nfp_net->dp,
1380                "control message budget exceeded!\n");
1381     }
1382 }
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!