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/overflow.h>
0007 #include <linux/sizes.h>
0008 #include <linux/bitfield.h>
0009 
0010 #include "../nfp_app.h"
0011 #include "../nfp_net.h"
0012 #include "../nfp_net_dp.h"
0013 #include "../crypto/crypto.h"
0014 #include "../crypto/fw.h"
0015 #include "nfdk.h"
0016 
0017 static int nfp_nfdk_tx_ring_should_wake(struct nfp_net_tx_ring *tx_ring)
0018 {
0019     return !nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT * 2);
0020 }
0021 
0022 static int nfp_nfdk_tx_ring_should_stop(struct nfp_net_tx_ring *tx_ring)
0023 {
0024     return nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT);
0025 }
0026 
0027 static void nfp_nfdk_tx_ring_stop(struct netdev_queue *nd_q,
0028                   struct nfp_net_tx_ring *tx_ring)
0029 {
0030     netif_tx_stop_queue(nd_q);
0031 
0032     /* We can race with the TX completion out of NAPI so recheck */
0033     smp_mb();
0034     if (unlikely(nfp_nfdk_tx_ring_should_wake(tx_ring)))
0035         netif_tx_start_queue(nd_q);
0036 }
0037 
0038 static __le64
0039 nfp_nfdk_tx_tso(struct nfp_net_r_vector *r_vec, struct nfp_nfdk_tx_buf *txbuf,
0040         struct sk_buff *skb)
0041 {
0042     u32 segs, hdrlen, l3_offset, l4_offset;
0043     struct nfp_nfdk_tx_desc txd;
0044     u16 mss;
0045 
0046     if (!skb->encapsulation) {
0047         l3_offset = skb_network_offset(skb);
0048         l4_offset = skb_transport_offset(skb);
0049         hdrlen = skb_tcp_all_headers(skb);
0050     } else {
0051         l3_offset = skb_inner_network_offset(skb);
0052         l4_offset = skb_inner_transport_offset(skb);
0053         hdrlen = skb_inner_tcp_all_headers(skb);
0054     }
0055 
0056     segs = skb_shinfo(skb)->gso_segs;
0057     mss = skb_shinfo(skb)->gso_size & NFDK_DESC_TX_MSS_MASK;
0058 
0059     txd.l3_offset = l3_offset;
0060     txd.l4_offset = l4_offset;
0061     txd.lso_meta_res = 0;
0062     txd.mss = cpu_to_le16(mss);
0063     txd.lso_hdrlen = hdrlen;
0064     txd.lso_totsegs = segs;
0065 
0066     txbuf->pkt_cnt = segs;
0067     txbuf->real_len = skb->len + hdrlen * (txbuf->pkt_cnt - 1);
0068 
0069     u64_stats_update_begin(&r_vec->tx_sync);
0070     r_vec->tx_lso++;
0071     u64_stats_update_end(&r_vec->tx_sync);
0072 
0073     return txd.raw;
0074 }
0075 
0076 static u8
0077 nfp_nfdk_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
0078          unsigned int pkt_cnt, struct sk_buff *skb, u64 flags)
0079 {
0080     struct ipv6hdr *ipv6h;
0081     struct iphdr *iph;
0082 
0083     if (!(dp->ctrl & NFP_NET_CFG_CTRL_TXCSUM))
0084         return flags;
0085 
0086     if (skb->ip_summed != CHECKSUM_PARTIAL)
0087         return flags;
0088 
0089     flags |= NFDK_DESC_TX_L4_CSUM;
0090 
0091     iph = skb->encapsulation ? inner_ip_hdr(skb) : ip_hdr(skb);
0092     ipv6h = skb->encapsulation ? inner_ipv6_hdr(skb) : ipv6_hdr(skb);
0093 
0094     /* L3 checksum offloading flag is not required for ipv6 */
0095     if (iph->version == 4) {
0096         flags |= NFDK_DESC_TX_L3_CSUM;
0097     } else if (ipv6h->version != 6) {
0098         nn_dp_warn(dp, "partial checksum but ipv=%x!\n", iph->version);
0099         return flags;
0100     }
0101 
0102     u64_stats_update_begin(&r_vec->tx_sync);
0103     if (!skb->encapsulation) {
0104         r_vec->hw_csum_tx += pkt_cnt;
0105     } else {
0106         flags |= NFDK_DESC_TX_ENCAP;
0107         r_vec->hw_csum_tx_inner += pkt_cnt;
0108     }
0109     u64_stats_update_end(&r_vec->tx_sync);
0110 
0111     return flags;
0112 }
0113 
0114 static int
0115 nfp_nfdk_tx_maybe_close_block(struct nfp_net_tx_ring *tx_ring,
0116                   struct sk_buff *skb)
0117 {
0118     unsigned int n_descs, wr_p, nop_slots;
0119     const skb_frag_t *frag, *fend;
0120     struct nfp_nfdk_tx_desc *txd;
0121     unsigned int nr_frags;
0122     unsigned int wr_idx;
0123     int err;
0124 
0125 recount_descs:
0126     n_descs = nfp_nfdk_headlen_to_segs(skb_headlen(skb));
0127     nr_frags = skb_shinfo(skb)->nr_frags;
0128     frag = skb_shinfo(skb)->frags;
0129     fend = frag + nr_frags;
0130     for (; frag < fend; frag++)
0131         n_descs += DIV_ROUND_UP(skb_frag_size(frag),
0132                     NFDK_TX_MAX_DATA_PER_DESC);
0133 
0134     if (unlikely(n_descs > NFDK_TX_DESC_GATHER_MAX)) {
0135         if (skb_is_nonlinear(skb)) {
0136             err = skb_linearize(skb);
0137             if (err)
0138                 return err;
0139             goto recount_descs;
0140         }
0141         return -EINVAL;
0142     }
0143 
0144     /* Under count by 1 (don't count meta) for the round down to work out */
0145     n_descs += !!skb_is_gso(skb);
0146 
0147     if (round_down(tx_ring->wr_p, NFDK_TX_DESC_BLOCK_CNT) !=
0148         round_down(tx_ring->wr_p + n_descs, NFDK_TX_DESC_BLOCK_CNT))
0149         goto close_block;
0150 
0151     if ((u32)tx_ring->data_pending + skb->len > NFDK_TX_MAX_DATA_PER_BLOCK)
0152         goto close_block;
0153 
0154     return 0;
0155 
0156 close_block:
0157     wr_p = tx_ring->wr_p;
0158     nop_slots = D_BLOCK_CPL(wr_p);
0159 
0160     wr_idx = D_IDX(tx_ring, wr_p);
0161     tx_ring->ktxbufs[wr_idx].skb = NULL;
0162     txd = &tx_ring->ktxds[wr_idx];
0163 
0164     memset(txd, 0, array_size(nop_slots, sizeof(struct nfp_nfdk_tx_desc)));
0165 
0166     tx_ring->data_pending = 0;
0167     tx_ring->wr_p += nop_slots;
0168     tx_ring->wr_ptr_add += nop_slots;
0169 
0170     return 0;
0171 }
0172 
0173 static int
0174 nfp_nfdk_prep_tx_meta(struct nfp_net_dp *dp, struct nfp_app *app,
0175               struct sk_buff *skb)
0176 {
0177     struct metadata_dst *md_dst = skb_metadata_dst(skb);
0178     unsigned char *data;
0179     bool vlan_insert;
0180     u32 meta_id = 0;
0181     int md_bytes;
0182 
0183     if (unlikely(md_dst && md_dst->type != METADATA_HW_PORT_MUX))
0184         md_dst = NULL;
0185 
0186     vlan_insert = skb_vlan_tag_present(skb) && (dp->ctrl & NFP_NET_CFG_CTRL_TXVLAN_V2);
0187 
0188     if (!(md_dst || vlan_insert))
0189         return 0;
0190 
0191     md_bytes = sizeof(meta_id) +
0192            !!md_dst * NFP_NET_META_PORTID_SIZE +
0193            vlan_insert * NFP_NET_META_VLAN_SIZE;
0194 
0195     if (unlikely(skb_cow_head(skb, md_bytes)))
0196         return -ENOMEM;
0197 
0198     data = skb_push(skb, md_bytes) + md_bytes;
0199     if (md_dst) {
0200         data -= NFP_NET_META_PORTID_SIZE;
0201         put_unaligned_be32(md_dst->u.port_info.port_id, data);
0202         meta_id = NFP_NET_META_PORTID;
0203     }
0204     if (vlan_insert) {
0205         data -= NFP_NET_META_VLAN_SIZE;
0206         /* data type of skb->vlan_proto is __be16
0207          * so it fills metadata without calling put_unaligned_be16
0208          */
0209         memcpy(data, &skb->vlan_proto, sizeof(skb->vlan_proto));
0210         put_unaligned_be16(skb_vlan_tag_get(skb), data + sizeof(skb->vlan_proto));
0211         meta_id <<= NFP_NET_META_FIELD_SIZE;
0212         meta_id |= NFP_NET_META_VLAN;
0213     }
0214 
0215     meta_id = FIELD_PREP(NFDK_META_LEN, md_bytes) |
0216           FIELD_PREP(NFDK_META_FIELDS, meta_id);
0217 
0218     data -= sizeof(meta_id);
0219     put_unaligned_be32(meta_id, data);
0220 
0221     return NFDK_DESC_TX_CHAIN_META;
0222 }
0223 
0224 /**
0225  * nfp_nfdk_tx() - Main transmit entry point
0226  * @skb:    SKB to transmit
0227  * @netdev: netdev structure
0228  *
0229  * Return: NETDEV_TX_OK on success.
0230  */
0231 netdev_tx_t nfp_nfdk_tx(struct sk_buff *skb, struct net_device *netdev)
0232 {
0233     struct nfp_net *nn = netdev_priv(netdev);
0234     struct nfp_nfdk_tx_buf *txbuf, *etxbuf;
0235     u32 cnt, tmp_dlen, dlen_type = 0;
0236     struct nfp_net_tx_ring *tx_ring;
0237     struct nfp_net_r_vector *r_vec;
0238     const skb_frag_t *frag, *fend;
0239     struct nfp_nfdk_tx_desc *txd;
0240     unsigned int real_len, qidx;
0241     unsigned int dma_len, type;
0242     struct netdev_queue *nd_q;
0243     struct nfp_net_dp *dp;
0244     int nr_frags, wr_idx;
0245     dma_addr_t dma_addr;
0246     u64 metadata;
0247 
0248     dp = &nn->dp;
0249     qidx = skb_get_queue_mapping(skb);
0250     tx_ring = &dp->tx_rings[qidx];
0251     r_vec = tx_ring->r_vec;
0252     nd_q = netdev_get_tx_queue(dp->netdev, qidx);
0253 
0254     /* Don't bother counting frags, assume the worst */
0255     if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
0256         nn_dp_warn(dp, "TX ring %d busy. wrp=%u rdp=%u\n",
0257                qidx, tx_ring->wr_p, tx_ring->rd_p);
0258         netif_tx_stop_queue(nd_q);
0259         nfp_net_tx_xmit_more_flush(tx_ring);
0260         u64_stats_update_begin(&r_vec->tx_sync);
0261         r_vec->tx_busy++;
0262         u64_stats_update_end(&r_vec->tx_sync);
0263         return NETDEV_TX_BUSY;
0264     }
0265 
0266     metadata = nfp_nfdk_prep_tx_meta(dp, nn->app, skb);
0267     if (unlikely((int)metadata < 0))
0268         goto err_flush;
0269 
0270     if (nfp_nfdk_tx_maybe_close_block(tx_ring, skb))
0271         goto err_flush;
0272 
0273     /* nr_frags will change after skb_linearize so we get nr_frags after
0274      * nfp_nfdk_tx_maybe_close_block function
0275      */
0276     nr_frags = skb_shinfo(skb)->nr_frags;
0277     /* DMA map all */
0278     wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
0279     txd = &tx_ring->ktxds[wr_idx];
0280     txbuf = &tx_ring->ktxbufs[wr_idx];
0281 
0282     dma_len = skb_headlen(skb);
0283     if (skb_is_gso(skb))
0284         type = NFDK_DESC_TX_TYPE_TSO;
0285     else if (!nr_frags && dma_len < NFDK_TX_MAX_DATA_PER_HEAD)
0286         type = NFDK_DESC_TX_TYPE_SIMPLE;
0287     else
0288         type = NFDK_DESC_TX_TYPE_GATHER;
0289 
0290     dma_addr = dma_map_single(dp->dev, skb->data, dma_len, DMA_TO_DEVICE);
0291     if (dma_mapping_error(dp->dev, dma_addr))
0292         goto err_warn_dma;
0293 
0294     txbuf->skb = skb;
0295     txbuf++;
0296 
0297     txbuf->dma_addr = dma_addr;
0298     txbuf++;
0299 
0300     /* FIELD_PREP() implicitly truncates to chunk */
0301     dma_len -= 1;
0302 
0303     /* We will do our best to pass as much data as we can in descriptor
0304      * and we need to make sure the first descriptor includes whole head
0305      * since there is limitation in firmware side. Sometimes the value of
0306      * dma_len bitwise and NFDK_DESC_TX_DMA_LEN_HEAD will less than
0307      * headlen.
0308      */
0309     dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
0310                    dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
0311                    NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
0312             FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
0313 
0314     txd->dma_len_type = cpu_to_le16(dlen_type);
0315     nfp_desc_set_dma_addr_48b(txd, dma_addr);
0316 
0317     /* starts at bit 0 */
0318     BUILD_BUG_ON(!(NFDK_DESC_TX_DMA_LEN_HEAD & 1));
0319 
0320     /* Preserve the original dlen_type, this way below the EOP logic
0321      * can use dlen_type.
0322      */
0323     tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
0324     dma_len -= tmp_dlen;
0325     dma_addr += tmp_dlen + 1;
0326     txd++;
0327 
0328     /* The rest of the data (if any) will be in larger dma descritors
0329      * and is handled with the fragment loop.
0330      */
0331     frag = skb_shinfo(skb)->frags;
0332     fend = frag + nr_frags;
0333 
0334     while (true) {
0335         while (dma_len > 0) {
0336             dma_len -= 1;
0337             dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN, dma_len);
0338 
0339             txd->dma_len_type = cpu_to_le16(dlen_type);
0340             nfp_desc_set_dma_addr_48b(txd, dma_addr);
0341 
0342             dma_len -= dlen_type;
0343             dma_addr += dlen_type + 1;
0344             txd++;
0345         }
0346 
0347         if (frag >= fend)
0348             break;
0349 
0350         dma_len = skb_frag_size(frag);
0351         dma_addr = skb_frag_dma_map(dp->dev, frag, 0, dma_len,
0352                         DMA_TO_DEVICE);
0353         if (dma_mapping_error(dp->dev, dma_addr))
0354             goto err_unmap;
0355 
0356         txbuf->dma_addr = dma_addr;
0357         txbuf++;
0358 
0359         frag++;
0360     }
0361 
0362     (txd - 1)->dma_len_type = cpu_to_le16(dlen_type | NFDK_DESC_TX_EOP);
0363 
0364     if (!skb_is_gso(skb)) {
0365         real_len = skb->len;
0366         /* Metadata desc */
0367         metadata = nfp_nfdk_tx_csum(dp, r_vec, 1, skb, metadata);
0368         txd->raw = cpu_to_le64(metadata);
0369         txd++;
0370     } else {
0371         /* lso desc should be placed after metadata desc */
0372         (txd + 1)->raw = nfp_nfdk_tx_tso(r_vec, txbuf, skb);
0373         real_len = txbuf->real_len;
0374         /* Metadata desc */
0375         metadata = nfp_nfdk_tx_csum(dp, r_vec, txbuf->pkt_cnt, skb, metadata);
0376         txd->raw = cpu_to_le64(metadata);
0377         txd += 2;
0378         txbuf++;
0379     }
0380 
0381     cnt = txd - tx_ring->ktxds - wr_idx;
0382     if (unlikely(round_down(wr_idx, NFDK_TX_DESC_BLOCK_CNT) !=
0383              round_down(wr_idx + cnt - 1, NFDK_TX_DESC_BLOCK_CNT)))
0384         goto err_warn_overflow;
0385 
0386     skb_tx_timestamp(skb);
0387 
0388     tx_ring->wr_p += cnt;
0389     if (tx_ring->wr_p % NFDK_TX_DESC_BLOCK_CNT)
0390         tx_ring->data_pending += skb->len;
0391     else
0392         tx_ring->data_pending = 0;
0393 
0394     if (nfp_nfdk_tx_ring_should_stop(tx_ring))
0395         nfp_nfdk_tx_ring_stop(nd_q, tx_ring);
0396 
0397     tx_ring->wr_ptr_add += cnt;
0398     if (__netdev_tx_sent_queue(nd_q, real_len, netdev_xmit_more()))
0399         nfp_net_tx_xmit_more_flush(tx_ring);
0400 
0401     return NETDEV_TX_OK;
0402 
0403 err_warn_overflow:
0404     WARN_ONCE(1, "unable to fit packet into a descriptor wr_idx:%d head:%d frags:%d cnt:%d",
0405           wr_idx, skb_headlen(skb), nr_frags, cnt);
0406     if (skb_is_gso(skb))
0407         txbuf--;
0408 err_unmap:
0409     /* txbuf pointed to the next-to-use */
0410     etxbuf = txbuf;
0411     /* first txbuf holds the skb */
0412     txbuf = &tx_ring->ktxbufs[wr_idx + 1];
0413     if (txbuf < etxbuf) {
0414         dma_unmap_single(dp->dev, txbuf->dma_addr,
0415                  skb_headlen(skb), DMA_TO_DEVICE);
0416         txbuf->raw = 0;
0417         txbuf++;
0418     }
0419     frag = skb_shinfo(skb)->frags;
0420     while (etxbuf < txbuf) {
0421         dma_unmap_page(dp->dev, txbuf->dma_addr,
0422                    skb_frag_size(frag), DMA_TO_DEVICE);
0423         txbuf->raw = 0;
0424         frag++;
0425         txbuf++;
0426     }
0427 err_warn_dma:
0428     nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
0429 err_flush:
0430     nfp_net_tx_xmit_more_flush(tx_ring);
0431     u64_stats_update_begin(&r_vec->tx_sync);
0432     r_vec->tx_errors++;
0433     u64_stats_update_end(&r_vec->tx_sync);
0434     dev_kfree_skb_any(skb);
0435     return NETDEV_TX_OK;
0436 }
0437 
0438 /**
0439  * nfp_nfdk_tx_complete() - Handled completed TX packets
0440  * @tx_ring:    TX ring structure
0441  * @budget: NAPI budget (only used as bool to determine if in NAPI context)
0442  */
0443 static void nfp_nfdk_tx_complete(struct nfp_net_tx_ring *tx_ring, int budget)
0444 {
0445     struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
0446     struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
0447     u32 done_pkts = 0, done_bytes = 0;
0448     struct nfp_nfdk_tx_buf *ktxbufs;
0449     struct device *dev = dp->dev;
0450     struct netdev_queue *nd_q;
0451     u32 rd_p, qcp_rd_p;
0452     int todo;
0453 
0454     rd_p = tx_ring->rd_p;
0455     if (tx_ring->wr_p == rd_p)
0456         return;
0457 
0458     /* Work out how many descriptors have been transmitted */
0459     qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
0460 
0461     if (qcp_rd_p == tx_ring->qcp_rd_p)
0462         return;
0463 
0464     todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
0465     ktxbufs = tx_ring->ktxbufs;
0466 
0467     while (todo > 0) {
0468         const skb_frag_t *frag, *fend;
0469         unsigned int size, n_descs = 1;
0470         struct nfp_nfdk_tx_buf *txbuf;
0471         struct sk_buff *skb;
0472 
0473         txbuf = &ktxbufs[D_IDX(tx_ring, rd_p)];
0474         skb = txbuf->skb;
0475         txbuf++;
0476 
0477         /* Closed block */
0478         if (!skb) {
0479             n_descs = D_BLOCK_CPL(rd_p);
0480             goto next;
0481         }
0482 
0483         /* Unmap head */
0484         size = skb_headlen(skb);
0485         n_descs += nfp_nfdk_headlen_to_segs(size);
0486         dma_unmap_single(dev, txbuf->dma_addr, size, DMA_TO_DEVICE);
0487         txbuf++;
0488 
0489         /* Unmap frags */
0490         frag = skb_shinfo(skb)->frags;
0491         fend = frag + skb_shinfo(skb)->nr_frags;
0492         for (; frag < fend; frag++) {
0493             size = skb_frag_size(frag);
0494             n_descs += DIV_ROUND_UP(size,
0495                         NFDK_TX_MAX_DATA_PER_DESC);
0496             dma_unmap_page(dev, txbuf->dma_addr,
0497                        skb_frag_size(frag), DMA_TO_DEVICE);
0498             txbuf++;
0499         }
0500 
0501         if (!skb_is_gso(skb)) {
0502             done_bytes += skb->len;
0503             done_pkts++;
0504         } else {
0505             done_bytes += txbuf->real_len;
0506             done_pkts += txbuf->pkt_cnt;
0507             n_descs++;
0508         }
0509 
0510         napi_consume_skb(skb, budget);
0511 next:
0512         rd_p += n_descs;
0513         todo -= n_descs;
0514     }
0515 
0516     tx_ring->rd_p = rd_p;
0517     tx_ring->qcp_rd_p = qcp_rd_p;
0518 
0519     u64_stats_update_begin(&r_vec->tx_sync);
0520     r_vec->tx_bytes += done_bytes;
0521     r_vec->tx_pkts += done_pkts;
0522     u64_stats_update_end(&r_vec->tx_sync);
0523 
0524     if (!dp->netdev)
0525         return;
0526 
0527     nd_q = netdev_get_tx_queue(dp->netdev, tx_ring->idx);
0528     netdev_tx_completed_queue(nd_q, done_pkts, done_bytes);
0529     if (nfp_nfdk_tx_ring_should_wake(tx_ring)) {
0530         /* Make sure TX thread will see updated tx_ring->rd_p */
0531         smp_mb();
0532 
0533         if (unlikely(netif_tx_queue_stopped(nd_q)))
0534             netif_tx_wake_queue(nd_q);
0535     }
0536 
0537     WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
0538           "TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
0539           tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
0540 }
0541 
0542 /* Receive processing */
0543 static void *
0544 nfp_nfdk_napi_alloc_one(struct nfp_net_dp *dp, dma_addr_t *dma_addr)
0545 {
0546     void *frag;
0547 
0548     if (!dp->xdp_prog) {
0549         frag = napi_alloc_frag(dp->fl_bufsz);
0550         if (unlikely(!frag))
0551             return NULL;
0552     } else {
0553         struct page *page;
0554 
0555         page = dev_alloc_page();
0556         if (unlikely(!page))
0557             return NULL;
0558         frag = page_address(page);
0559     }
0560 
0561     *dma_addr = nfp_net_dma_map_rx(dp, frag);
0562     if (dma_mapping_error(dp->dev, *dma_addr)) {
0563         nfp_net_free_frag(frag, dp->xdp_prog);
0564         nn_dp_warn(dp, "Failed to map DMA RX buffer\n");
0565         return NULL;
0566     }
0567 
0568     return frag;
0569 }
0570 
0571 /**
0572  * nfp_nfdk_rx_give_one() - Put mapped skb on the software and hardware rings
0573  * @dp:     NFP Net data path struct
0574  * @rx_ring:    RX ring structure
0575  * @frag:   page fragment buffer
0576  * @dma_addr:   DMA address of skb mapping
0577  */
0578 static void
0579 nfp_nfdk_rx_give_one(const struct nfp_net_dp *dp,
0580              struct nfp_net_rx_ring *rx_ring,
0581              void *frag, dma_addr_t dma_addr)
0582 {
0583     unsigned int wr_idx;
0584 
0585     wr_idx = D_IDX(rx_ring, rx_ring->wr_p);
0586 
0587     nfp_net_dma_sync_dev_rx(dp, dma_addr);
0588 
0589     /* Stash SKB and DMA address away */
0590     rx_ring->rxbufs[wr_idx].frag = frag;
0591     rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
0592 
0593     /* Fill freelist descriptor */
0594     rx_ring->rxds[wr_idx].fld.reserved = 0;
0595     rx_ring->rxds[wr_idx].fld.meta_len_dd = 0;
0596     nfp_desc_set_dma_addr_48b(&rx_ring->rxds[wr_idx].fld,
0597                   dma_addr + dp->rx_dma_off);
0598 
0599     rx_ring->wr_p++;
0600     if (!(rx_ring->wr_p % NFP_NET_FL_BATCH)) {
0601         /* Update write pointer of the freelist queue. Make
0602          * sure all writes are flushed before telling the hardware.
0603          */
0604         wmb();
0605         nfp_qcp_wr_ptr_add(rx_ring->qcp_fl, NFP_NET_FL_BATCH);
0606     }
0607 }
0608 
0609 /**
0610  * nfp_nfdk_rx_ring_fill_freelist() - Give buffers from the ring to FW
0611  * @dp:      NFP Net data path struct
0612  * @rx_ring: RX ring to fill
0613  */
0614 void nfp_nfdk_rx_ring_fill_freelist(struct nfp_net_dp *dp,
0615                     struct nfp_net_rx_ring *rx_ring)
0616 {
0617     unsigned int i;
0618 
0619     for (i = 0; i < rx_ring->cnt - 1; i++)
0620         nfp_nfdk_rx_give_one(dp, rx_ring, rx_ring->rxbufs[i].frag,
0621                      rx_ring->rxbufs[i].dma_addr);
0622 }
0623 
0624 /**
0625  * nfp_nfdk_rx_csum_has_errors() - group check if rxd has any csum errors
0626  * @flags: RX descriptor flags field in CPU byte order
0627  */
0628 static int nfp_nfdk_rx_csum_has_errors(u16 flags)
0629 {
0630     u16 csum_all_checked, csum_all_ok;
0631 
0632     csum_all_checked = flags & __PCIE_DESC_RX_CSUM_ALL;
0633     csum_all_ok = flags & __PCIE_DESC_RX_CSUM_ALL_OK;
0634 
0635     return csum_all_checked != (csum_all_ok << PCIE_DESC_RX_CSUM_OK_SHIFT);
0636 }
0637 
0638 /**
0639  * nfp_nfdk_rx_csum() - set SKB checksum field based on RX descriptor flags
0640  * @dp:  NFP Net data path struct
0641  * @r_vec: per-ring structure
0642  * @rxd: Pointer to RX descriptor
0643  * @meta: Parsed metadata prepend
0644  * @skb: Pointer to SKB
0645  */
0646 static void
0647 nfp_nfdk_rx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
0648          struct nfp_net_rx_desc *rxd, struct nfp_meta_parsed *meta,
0649          struct sk_buff *skb)
0650 {
0651     skb_checksum_none_assert(skb);
0652 
0653     if (!(dp->netdev->features & NETIF_F_RXCSUM))
0654         return;
0655 
0656     if (meta->csum_type) {
0657         skb->ip_summed = meta->csum_type;
0658         skb->csum = meta->csum;
0659         u64_stats_update_begin(&r_vec->rx_sync);
0660         r_vec->hw_csum_rx_complete++;
0661         u64_stats_update_end(&r_vec->rx_sync);
0662         return;
0663     }
0664 
0665     if (nfp_nfdk_rx_csum_has_errors(le16_to_cpu(rxd->rxd.flags))) {
0666         u64_stats_update_begin(&r_vec->rx_sync);
0667         r_vec->hw_csum_rx_error++;
0668         u64_stats_update_end(&r_vec->rx_sync);
0669         return;
0670     }
0671 
0672     /* Assume that the firmware will never report inner CSUM_OK unless outer
0673      * L4 headers were successfully parsed. FW will always report zero UDP
0674      * checksum as CSUM_OK.
0675      */
0676     if (rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK ||
0677         rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK) {
0678         __skb_incr_checksum_unnecessary(skb);
0679         u64_stats_update_begin(&r_vec->rx_sync);
0680         r_vec->hw_csum_rx_ok++;
0681         u64_stats_update_end(&r_vec->rx_sync);
0682     }
0683 
0684     if (rxd->rxd.flags & PCIE_DESC_RX_I_TCP_CSUM_OK ||
0685         rxd->rxd.flags & PCIE_DESC_RX_I_UDP_CSUM_OK) {
0686         __skb_incr_checksum_unnecessary(skb);
0687         u64_stats_update_begin(&r_vec->rx_sync);
0688         r_vec->hw_csum_rx_inner_ok++;
0689         u64_stats_update_end(&r_vec->rx_sync);
0690     }
0691 }
0692 
0693 static void
0694 nfp_nfdk_set_hash(struct net_device *netdev, struct nfp_meta_parsed *meta,
0695           unsigned int type, __be32 *hash)
0696 {
0697     if (!(netdev->features & NETIF_F_RXHASH))
0698         return;
0699 
0700     switch (type) {
0701     case NFP_NET_RSS_IPV4:
0702     case NFP_NET_RSS_IPV6:
0703     case NFP_NET_RSS_IPV6_EX:
0704         meta->hash_type = PKT_HASH_TYPE_L3;
0705         break;
0706     default:
0707         meta->hash_type = PKT_HASH_TYPE_L4;
0708         break;
0709     }
0710 
0711     meta->hash = get_unaligned_be32(hash);
0712 }
0713 
0714 static bool
0715 nfp_nfdk_parse_meta(struct net_device *netdev, struct nfp_meta_parsed *meta,
0716             void *data, void *pkt, unsigned int pkt_len, int meta_len)
0717 {
0718     u32 meta_info, vlan_info;
0719 
0720     meta_info = get_unaligned_be32(data);
0721     data += 4;
0722 
0723     while (meta_info) {
0724         switch (meta_info & NFP_NET_META_FIELD_MASK) {
0725         case NFP_NET_META_HASH:
0726             meta_info >>= NFP_NET_META_FIELD_SIZE;
0727             nfp_nfdk_set_hash(netdev, meta,
0728                       meta_info & NFP_NET_META_FIELD_MASK,
0729                       (__be32 *)data);
0730             data += 4;
0731             break;
0732         case NFP_NET_META_MARK:
0733             meta->mark = get_unaligned_be32(data);
0734             data += 4;
0735             break;
0736         case NFP_NET_META_VLAN:
0737             vlan_info = get_unaligned_be32(data);
0738             if (FIELD_GET(NFP_NET_META_VLAN_STRIP, vlan_info)) {
0739                 meta->vlan.stripped = true;
0740                 meta->vlan.tpid = FIELD_GET(NFP_NET_META_VLAN_TPID_MASK,
0741                                 vlan_info);
0742                 meta->vlan.tci = FIELD_GET(NFP_NET_META_VLAN_TCI_MASK,
0743                                vlan_info);
0744             }
0745             data += 4;
0746             break;
0747         case NFP_NET_META_PORTID:
0748             meta->portid = get_unaligned_be32(data);
0749             data += 4;
0750             break;
0751         case NFP_NET_META_CSUM:
0752             meta->csum_type = CHECKSUM_COMPLETE;
0753             meta->csum =
0754                 (__force __wsum)__get_unaligned_cpu32(data);
0755             data += 4;
0756             break;
0757         case NFP_NET_META_RESYNC_INFO:
0758             if (nfp_net_tls_rx_resync_req(netdev, data, pkt,
0759                               pkt_len))
0760                 return false;
0761             data += sizeof(struct nfp_net_tls_resync_req);
0762             break;
0763         default:
0764             return true;
0765         }
0766 
0767         meta_info >>= NFP_NET_META_FIELD_SIZE;
0768     }
0769 
0770     return data != pkt;
0771 }
0772 
0773 static void
0774 nfp_nfdk_rx_drop(const struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
0775          struct nfp_net_rx_ring *rx_ring, struct nfp_net_rx_buf *rxbuf,
0776          struct sk_buff *skb)
0777 {
0778     u64_stats_update_begin(&r_vec->rx_sync);
0779     r_vec->rx_drops++;
0780     /* If we have both skb and rxbuf the replacement buffer allocation
0781      * must have failed, count this as an alloc failure.
0782      */
0783     if (skb && rxbuf)
0784         r_vec->rx_replace_buf_alloc_fail++;
0785     u64_stats_update_end(&r_vec->rx_sync);
0786 
0787     /* skb is build based on the frag, free_skb() would free the frag
0788      * so to be able to reuse it we need an extra ref.
0789      */
0790     if (skb && rxbuf && skb->head == rxbuf->frag)
0791         page_ref_inc(virt_to_head_page(rxbuf->frag));
0792     if (rxbuf)
0793         nfp_nfdk_rx_give_one(dp, rx_ring, rxbuf->frag, rxbuf->dma_addr);
0794     if (skb)
0795         dev_kfree_skb_any(skb);
0796 }
0797 
0798 static bool nfp_nfdk_xdp_complete(struct nfp_net_tx_ring *tx_ring)
0799 {
0800     struct nfp_net_r_vector *r_vec = tx_ring->r_vec;
0801     struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
0802     struct nfp_net_rx_ring *rx_ring;
0803     u32 qcp_rd_p, done = 0;
0804     bool done_all;
0805     int todo;
0806 
0807     /* Work out how many descriptors have been transmitted */
0808     qcp_rd_p = nfp_net_read_tx_cmpl(tx_ring, dp);
0809     if (qcp_rd_p == tx_ring->qcp_rd_p)
0810         return true;
0811 
0812     todo = D_IDX(tx_ring, qcp_rd_p - tx_ring->qcp_rd_p);
0813 
0814     done_all = todo <= NFP_NET_XDP_MAX_COMPLETE;
0815     todo = min(todo, NFP_NET_XDP_MAX_COMPLETE);
0816 
0817     rx_ring = r_vec->rx_ring;
0818     while (todo > 0) {
0819         int idx = D_IDX(tx_ring, tx_ring->rd_p + done);
0820         struct nfp_nfdk_tx_buf *txbuf;
0821         unsigned int step = 1;
0822 
0823         txbuf = &tx_ring->ktxbufs[idx];
0824         if (!txbuf->raw)
0825             goto next;
0826 
0827         if (NFDK_TX_BUF_INFO(txbuf->val) != NFDK_TX_BUF_INFO_SOP) {
0828             WARN_ONCE(1, "Unexpected TX buffer in XDP TX ring\n");
0829             goto next;
0830         }
0831 
0832         /* Two successive txbufs are used to stash virtual and dma
0833          * address respectively, recycle and clean them here.
0834          */
0835         nfp_nfdk_rx_give_one(dp, rx_ring,
0836                      (void *)NFDK_TX_BUF_PTR(txbuf[0].val),
0837                      txbuf[1].dma_addr);
0838         txbuf[0].raw = 0;
0839         txbuf[1].raw = 0;
0840         step = 2;
0841 
0842         u64_stats_update_begin(&r_vec->tx_sync);
0843         /* Note: tx_bytes not accumulated. */
0844         r_vec->tx_pkts++;
0845         u64_stats_update_end(&r_vec->tx_sync);
0846 next:
0847         todo -= step;
0848         done += step;
0849     }
0850 
0851     tx_ring->qcp_rd_p = D_IDX(tx_ring, tx_ring->qcp_rd_p + done);
0852     tx_ring->rd_p += done;
0853 
0854     WARN_ONCE(tx_ring->wr_p - tx_ring->rd_p > tx_ring->cnt,
0855           "XDP TX ring corruption rd_p=%u wr_p=%u cnt=%u\n",
0856           tx_ring->rd_p, tx_ring->wr_p, tx_ring->cnt);
0857 
0858     return done_all;
0859 }
0860 
0861 static bool
0862 nfp_nfdk_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
0863             struct nfp_net_tx_ring *tx_ring,
0864             struct nfp_net_rx_buf *rxbuf, unsigned int dma_off,
0865             unsigned int pkt_len, bool *completed)
0866 {
0867     unsigned int dma_map_sz = dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA;
0868     unsigned int dma_len, type, cnt, dlen_type, tmp_dlen;
0869     struct nfp_nfdk_tx_buf *txbuf;
0870     struct nfp_nfdk_tx_desc *txd;
0871     unsigned int n_descs;
0872     dma_addr_t dma_addr;
0873     int wr_idx;
0874 
0875     /* Reject if xdp_adjust_tail grow packet beyond DMA area */
0876     if (pkt_len + dma_off > dma_map_sz)
0877         return false;
0878 
0879     /* Make sure there's still at least one block available after
0880      * aligning to block boundary, so that the txds used below
0881      * won't wrap around the tx_ring.
0882      */
0883     if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
0884         if (!*completed) {
0885             nfp_nfdk_xdp_complete(tx_ring);
0886             *completed = true;
0887         }
0888 
0889         if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
0890             nfp_nfdk_rx_drop(dp, rx_ring->r_vec, rx_ring, rxbuf,
0891                      NULL);
0892             return false;
0893         }
0894     }
0895 
0896     /* Check if cross block boundary */
0897     n_descs = nfp_nfdk_headlen_to_segs(pkt_len);
0898     if ((round_down(tx_ring->wr_p, NFDK_TX_DESC_BLOCK_CNT) !=
0899          round_down(tx_ring->wr_p + n_descs, NFDK_TX_DESC_BLOCK_CNT)) ||
0900         ((u32)tx_ring->data_pending + pkt_len >
0901          NFDK_TX_MAX_DATA_PER_BLOCK)) {
0902         unsigned int nop_slots = D_BLOCK_CPL(tx_ring->wr_p);
0903 
0904         wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
0905         txd = &tx_ring->ktxds[wr_idx];
0906         memset(txd, 0,
0907                array_size(nop_slots, sizeof(struct nfp_nfdk_tx_desc)));
0908 
0909         tx_ring->data_pending = 0;
0910         tx_ring->wr_p += nop_slots;
0911         tx_ring->wr_ptr_add += nop_slots;
0912     }
0913 
0914     wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
0915 
0916     txbuf = &tx_ring->ktxbufs[wr_idx];
0917 
0918     txbuf[0].val = (unsigned long)rxbuf->frag | NFDK_TX_BUF_INFO_SOP;
0919     txbuf[1].dma_addr = rxbuf->dma_addr;
0920     /* Note: pkt len not stored */
0921 
0922     dma_sync_single_for_device(dp->dev, rxbuf->dma_addr + dma_off,
0923                    pkt_len, DMA_BIDIRECTIONAL);
0924 
0925     /* Build TX descriptor */
0926     txd = &tx_ring->ktxds[wr_idx];
0927     dma_len = pkt_len;
0928     dma_addr = rxbuf->dma_addr + dma_off;
0929 
0930     if (dma_len < NFDK_TX_MAX_DATA_PER_HEAD)
0931         type = NFDK_DESC_TX_TYPE_SIMPLE;
0932     else
0933         type = NFDK_DESC_TX_TYPE_GATHER;
0934 
0935     /* FIELD_PREP() implicitly truncates to chunk */
0936     dma_len -= 1;
0937     dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
0938                    dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
0939                    NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
0940             FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
0941 
0942     txd->dma_len_type = cpu_to_le16(dlen_type);
0943     nfp_desc_set_dma_addr_48b(txd, dma_addr);
0944 
0945     tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
0946     dma_len -= tmp_dlen;
0947     dma_addr += tmp_dlen + 1;
0948     txd++;
0949 
0950     while (dma_len > 0) {
0951         dma_len -= 1;
0952         dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN, dma_len);
0953         txd->dma_len_type = cpu_to_le16(dlen_type);
0954         nfp_desc_set_dma_addr_48b(txd, dma_addr);
0955 
0956         dlen_type &= NFDK_DESC_TX_DMA_LEN;
0957         dma_len -= dlen_type;
0958         dma_addr += dlen_type + 1;
0959         txd++;
0960     }
0961 
0962     (txd - 1)->dma_len_type = cpu_to_le16(dlen_type | NFDK_DESC_TX_EOP);
0963 
0964     /* Metadata desc */
0965     txd->raw = 0;
0966     txd++;
0967 
0968     cnt = txd - tx_ring->ktxds - wr_idx;
0969     tx_ring->wr_p += cnt;
0970     if (tx_ring->wr_p % NFDK_TX_DESC_BLOCK_CNT)
0971         tx_ring->data_pending += pkt_len;
0972     else
0973         tx_ring->data_pending = 0;
0974 
0975     tx_ring->wr_ptr_add += cnt;
0976     return true;
0977 }
0978 
0979 /**
0980  * nfp_nfdk_rx() - receive up to @budget packets on @rx_ring
0981  * @rx_ring:   RX ring to receive from
0982  * @budget:    NAPI budget
0983  *
0984  * Note, this function is separated out from the napi poll function to
0985  * more cleanly separate packet receive code from other bookkeeping
0986  * functions performed in the napi poll function.
0987  *
0988  * Return: Number of packets received.
0989  */
0990 static int nfp_nfdk_rx(struct nfp_net_rx_ring *rx_ring, int budget)
0991 {
0992     struct nfp_net_r_vector *r_vec = rx_ring->r_vec;
0993     struct nfp_net_dp *dp = &r_vec->nfp_net->dp;
0994     struct nfp_net_tx_ring *tx_ring;
0995     struct bpf_prog *xdp_prog;
0996     bool xdp_tx_cmpl = false;
0997     unsigned int true_bufsz;
0998     struct sk_buff *skb;
0999     int pkts_polled = 0;
1000     struct xdp_buff xdp;
1001     int idx;
1002 
1003     xdp_prog = READ_ONCE(dp->xdp_prog);
1004     true_bufsz = xdp_prog ? PAGE_SIZE : dp->fl_bufsz;
1005     xdp_init_buff(&xdp, PAGE_SIZE - NFP_NET_RX_BUF_HEADROOM,
1006               &rx_ring->xdp_rxq);
1007     tx_ring = r_vec->xdp_ring;
1008 
1009     while (pkts_polled < budget) {
1010         unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1011         struct nfp_net_rx_buf *rxbuf;
1012         struct nfp_net_rx_desc *rxd;
1013         struct nfp_meta_parsed meta;
1014         bool redir_egress = false;
1015         struct net_device *netdev;
1016         dma_addr_t new_dma_addr;
1017         u32 meta_len_xdp = 0;
1018         void *new_frag;
1019 
1020         idx = D_IDX(rx_ring, rx_ring->rd_p);
1021 
1022         rxd = &rx_ring->rxds[idx];
1023         if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1024             break;
1025 
1026         /* Memory barrier to ensure that we won't do other reads
1027          * before the DD bit.
1028          */
1029         dma_rmb();
1030 
1031         memset(&meta, 0, sizeof(meta));
1032 
1033         rx_ring->rd_p++;
1034         pkts_polled++;
1035 
1036         rxbuf = &rx_ring->rxbufs[idx];
1037         /*         < meta_len >
1038          *  <-- [rx_offset] -->
1039          *  ---------------------------------------------------------
1040          * | [XX] |  metadata  |             packet           | XXXX |
1041          *  ---------------------------------------------------------
1042          *         <---------------- data_len --------------->
1043          *
1044          * The rx_offset is fixed for all packets, the meta_len can vary
1045          * on a packet by packet basis. If rx_offset is set to zero
1046          * (_RX_OFFSET_DYNAMIC) metadata starts at the beginning of the
1047          * buffer and is immediately followed by the packet (no [XX]).
1048          */
1049         meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1050         data_len = le16_to_cpu(rxd->rxd.data_len);
1051         pkt_len = data_len - meta_len;
1052 
1053         pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1054         if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1055             pkt_off += meta_len;
1056         else
1057             pkt_off += dp->rx_offset;
1058         meta_off = pkt_off - meta_len;
1059 
1060         /* Stats update */
1061         u64_stats_update_begin(&r_vec->rx_sync);
1062         r_vec->rx_pkts++;
1063         r_vec->rx_bytes += pkt_len;
1064         u64_stats_update_end(&r_vec->rx_sync);
1065 
1066         if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
1067                  (dp->rx_offset && meta_len > dp->rx_offset))) {
1068             nn_dp_warn(dp, "oversized RX packet metadata %u\n",
1069                    meta_len);
1070             nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1071             continue;
1072         }
1073 
1074         nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
1075                     data_len);
1076 
1077         if (meta_len) {
1078             if (unlikely(nfp_nfdk_parse_meta(dp->netdev, &meta,
1079                              rxbuf->frag + meta_off,
1080                              rxbuf->frag + pkt_off,
1081                              pkt_len, meta_len))) {
1082                 nn_dp_warn(dp, "invalid RX packet metadata\n");
1083                 nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf,
1084                          NULL);
1085                 continue;
1086             }
1087         }
1088 
1089         if (xdp_prog && !meta.portid) {
1090             void *orig_data = rxbuf->frag + pkt_off;
1091             unsigned int dma_off;
1092             int act;
1093 
1094             xdp_prepare_buff(&xdp,
1095                      rxbuf->frag + NFP_NET_RX_BUF_HEADROOM,
1096                      pkt_off - NFP_NET_RX_BUF_HEADROOM,
1097                      pkt_len, true);
1098 
1099             act = bpf_prog_run_xdp(xdp_prog, &xdp);
1100 
1101             pkt_len = xdp.data_end - xdp.data;
1102             pkt_off += xdp.data - orig_data;
1103 
1104             switch (act) {
1105             case XDP_PASS:
1106                 meta_len_xdp = xdp.data - xdp.data_meta;
1107                 break;
1108             case XDP_TX:
1109                 dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
1110                 if (unlikely(!nfp_nfdk_tx_xdp_buf(dp, rx_ring,
1111                                   tx_ring,
1112                                   rxbuf,
1113                                   dma_off,
1114                                   pkt_len,
1115                                   &xdp_tx_cmpl)))
1116                     trace_xdp_exception(dp->netdev,
1117                                 xdp_prog, act);
1118                 continue;
1119             default:
1120                 bpf_warn_invalid_xdp_action(dp->netdev, xdp_prog, act);
1121                 fallthrough;
1122             case XDP_ABORTED:
1123                 trace_xdp_exception(dp->netdev, xdp_prog, act);
1124                 fallthrough;
1125             case XDP_DROP:
1126                 nfp_nfdk_rx_give_one(dp, rx_ring, rxbuf->frag,
1127                              rxbuf->dma_addr);
1128                 continue;
1129             }
1130         }
1131 
1132         if (likely(!meta.portid)) {
1133             netdev = dp->netdev;
1134         } else if (meta.portid == NFP_META_PORT_ID_CTRL) {
1135             struct nfp_net *nn = netdev_priv(dp->netdev);
1136 
1137             nfp_app_ctrl_rx_raw(nn->app, rxbuf->frag + pkt_off,
1138                         pkt_len);
1139             nfp_nfdk_rx_give_one(dp, rx_ring, rxbuf->frag,
1140                          rxbuf->dma_addr);
1141             continue;
1142         } else {
1143             struct nfp_net *nn;
1144 
1145             nn = netdev_priv(dp->netdev);
1146             netdev = nfp_app_dev_get(nn->app, meta.portid,
1147                          &redir_egress);
1148             if (unlikely(!netdev)) {
1149                 nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf,
1150                          NULL);
1151                 continue;
1152             }
1153 
1154             if (nfp_netdev_is_nfp_repr(netdev))
1155                 nfp_repr_inc_rx_stats(netdev, pkt_len);
1156         }
1157 
1158         skb = build_skb(rxbuf->frag, true_bufsz);
1159         if (unlikely(!skb)) {
1160             nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1161             continue;
1162         }
1163         new_frag = nfp_nfdk_napi_alloc_one(dp, &new_dma_addr);
1164         if (unlikely(!new_frag)) {
1165             nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1166             continue;
1167         }
1168 
1169         nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1170 
1171         nfp_nfdk_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1172 
1173         skb_reserve(skb, pkt_off);
1174         skb_put(skb, pkt_len);
1175 
1176         skb->mark = meta.mark;
1177         skb_set_hash(skb, meta.hash, meta.hash_type);
1178 
1179         skb_record_rx_queue(skb, rx_ring->idx);
1180         skb->protocol = eth_type_trans(skb, netdev);
1181 
1182         nfp_nfdk_rx_csum(dp, r_vec, rxd, &meta, skb);
1183 
1184         if (unlikely(!nfp_net_vlan_strip(skb, rxd, &meta))) {
1185             nfp_nfdk_rx_drop(dp, r_vec, rx_ring, NULL, skb);
1186             continue;
1187         }
1188 
1189         if (meta_len_xdp)
1190             skb_metadata_set(skb, meta_len_xdp);
1191 
1192         if (likely(!redir_egress)) {
1193             napi_gro_receive(&rx_ring->r_vec->napi, skb);
1194         } else {
1195             skb->dev = netdev;
1196             skb_reset_network_header(skb);
1197             __skb_push(skb, ETH_HLEN);
1198             dev_queue_xmit(skb);
1199         }
1200     }
1201 
1202     if (xdp_prog) {
1203         if (tx_ring->wr_ptr_add)
1204             nfp_net_tx_xmit_more_flush(tx_ring);
1205         else if (unlikely(tx_ring->wr_p != tx_ring->rd_p) &&
1206              !xdp_tx_cmpl)
1207             if (!nfp_nfdk_xdp_complete(tx_ring))
1208                 pkts_polled = budget;
1209     }
1210 
1211     return pkts_polled;
1212 }
1213 
1214 /**
1215  * nfp_nfdk_poll() - napi poll function
1216  * @napi:    NAPI structure
1217  * @budget:  NAPI budget
1218  *
1219  * Return: number of packets polled.
1220  */
1221 int nfp_nfdk_poll(struct napi_struct *napi, int budget)
1222 {
1223     struct nfp_net_r_vector *r_vec =
1224         container_of(napi, struct nfp_net_r_vector, napi);
1225     unsigned int pkts_polled = 0;
1226 
1227     if (r_vec->tx_ring)
1228         nfp_nfdk_tx_complete(r_vec->tx_ring, budget);
1229     if (r_vec->rx_ring)
1230         pkts_polled = nfp_nfdk_rx(r_vec->rx_ring, budget);
1231 
1232     if (pkts_polled < budget)
1233         if (napi_complete_done(napi, pkts_polled))
1234             nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1235 
1236     if (r_vec->nfp_net->rx_coalesce_adapt_on && r_vec->rx_ring) {
1237         struct dim_sample dim_sample = {};
1238         unsigned int start;
1239         u64 pkts, bytes;
1240 
1241         do {
1242             start = u64_stats_fetch_begin(&r_vec->rx_sync);
1243             pkts = r_vec->rx_pkts;
1244             bytes = r_vec->rx_bytes;
1245         } while (u64_stats_fetch_retry(&r_vec->rx_sync, start));
1246 
1247         dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1248         net_dim(&r_vec->rx_dim, dim_sample);
1249     }
1250 
1251     if (r_vec->nfp_net->tx_coalesce_adapt_on && r_vec->tx_ring) {
1252         struct dim_sample dim_sample = {};
1253         unsigned int start;
1254         u64 pkts, bytes;
1255 
1256         do {
1257             start = u64_stats_fetch_begin(&r_vec->tx_sync);
1258             pkts = r_vec->tx_pkts;
1259             bytes = r_vec->tx_bytes;
1260         } while (u64_stats_fetch_retry(&r_vec->tx_sync, start));
1261 
1262         dim_update_sample(r_vec->event_ctr, pkts, bytes, &dim_sample);
1263         net_dim(&r_vec->tx_dim, dim_sample);
1264     }
1265 
1266     return pkts_polled;
1267 }
1268 
1269 /* Control device data path
1270  */
1271 
1272 bool
1273 nfp_nfdk_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
1274              struct sk_buff *skb, bool old)
1275 {
1276     u32 cnt, tmp_dlen, dlen_type = 0;
1277     struct nfp_net_tx_ring *tx_ring;
1278     struct nfp_nfdk_tx_buf *txbuf;
1279     struct nfp_nfdk_tx_desc *txd;
1280     unsigned int dma_len, type;
1281     struct nfp_net_dp *dp;
1282     dma_addr_t dma_addr;
1283     u64 metadata = 0;
1284     int wr_idx;
1285 
1286     dp = &r_vec->nfp_net->dp;
1287     tx_ring = r_vec->tx_ring;
1288 
1289     if (WARN_ON_ONCE(skb_shinfo(skb)->nr_frags)) {
1290         nn_dp_warn(dp, "Driver's CTRL TX does not implement gather\n");
1291         goto err_free;
1292     }
1293 
1294     /* Don't bother counting frags, assume the worst */
1295     if (unlikely(nfp_net_tx_full(tx_ring, NFDK_TX_DESC_STOP_CNT))) {
1296         u64_stats_update_begin(&r_vec->tx_sync);
1297         r_vec->tx_busy++;
1298         u64_stats_update_end(&r_vec->tx_sync);
1299         if (!old)
1300             __skb_queue_tail(&r_vec->queue, skb);
1301         else
1302             __skb_queue_head(&r_vec->queue, skb);
1303         return NETDEV_TX_BUSY;
1304     }
1305 
1306     if (nfp_app_ctrl_has_meta(nn->app)) {
1307         if (unlikely(skb_headroom(skb) < 8)) {
1308             nn_dp_warn(dp, "CTRL TX on skb without headroom\n");
1309             goto err_free;
1310         }
1311         metadata = NFDK_DESC_TX_CHAIN_META;
1312         put_unaligned_be32(NFP_META_PORT_ID_CTRL, skb_push(skb, 4));
1313         put_unaligned_be32(FIELD_PREP(NFDK_META_LEN, 8) |
1314                    FIELD_PREP(NFDK_META_FIELDS,
1315                           NFP_NET_META_PORTID),
1316                    skb_push(skb, 4));
1317     }
1318 
1319     if (nfp_nfdk_tx_maybe_close_block(tx_ring, skb))
1320         goto err_free;
1321 
1322     /* DMA map all */
1323     wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
1324     txd = &tx_ring->ktxds[wr_idx];
1325     txbuf = &tx_ring->ktxbufs[wr_idx];
1326 
1327     dma_len = skb_headlen(skb);
1328     if (dma_len < NFDK_TX_MAX_DATA_PER_HEAD)
1329         type = NFDK_DESC_TX_TYPE_SIMPLE;
1330     else
1331         type = NFDK_DESC_TX_TYPE_GATHER;
1332 
1333     dma_addr = dma_map_single(dp->dev, skb->data, dma_len, DMA_TO_DEVICE);
1334     if (dma_mapping_error(dp->dev, dma_addr))
1335         goto err_warn_dma;
1336 
1337     txbuf->skb = skb;
1338     txbuf++;
1339 
1340     txbuf->dma_addr = dma_addr;
1341     txbuf++;
1342 
1343     dma_len -= 1;
1344     dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
1345                    dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
1346                    NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
1347             FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
1348 
1349     txd->dma_len_type = cpu_to_le16(dlen_type);
1350     nfp_desc_set_dma_addr_48b(txd, dma_addr);
1351 
1352     tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
1353     dma_len -= tmp_dlen;
1354     dma_addr += tmp_dlen + 1;
1355     txd++;
1356 
1357     while (dma_len > 0) {
1358         dma_len -= 1;
1359         dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN, dma_len);
1360         txd->dma_len_type = cpu_to_le16(dlen_type);
1361         nfp_desc_set_dma_addr_48b(txd, dma_addr);
1362 
1363         dlen_type &= NFDK_DESC_TX_DMA_LEN;
1364         dma_len -= dlen_type;
1365         dma_addr += dlen_type + 1;
1366         txd++;
1367     }
1368 
1369     (txd - 1)->dma_len_type = cpu_to_le16(dlen_type | NFDK_DESC_TX_EOP);
1370 
1371     /* Metadata desc */
1372     txd->raw = cpu_to_le64(metadata);
1373     txd++;
1374 
1375     cnt = txd - tx_ring->ktxds - wr_idx;
1376     if (unlikely(round_down(wr_idx, NFDK_TX_DESC_BLOCK_CNT) !=
1377              round_down(wr_idx + cnt - 1, NFDK_TX_DESC_BLOCK_CNT)))
1378         goto err_warn_overflow;
1379 
1380     tx_ring->wr_p += cnt;
1381     if (tx_ring->wr_p % NFDK_TX_DESC_BLOCK_CNT)
1382         tx_ring->data_pending += skb->len;
1383     else
1384         tx_ring->data_pending = 0;
1385 
1386     tx_ring->wr_ptr_add += cnt;
1387     nfp_net_tx_xmit_more_flush(tx_ring);
1388 
1389     return NETDEV_TX_OK;
1390 
1391 err_warn_overflow:
1392     WARN_ONCE(1, "unable to fit packet into a descriptor wr_idx:%d head:%d frags:%d cnt:%d",
1393           wr_idx, skb_headlen(skb), 0, cnt);
1394     txbuf--;
1395     dma_unmap_single(dp->dev, txbuf->dma_addr,
1396              skb_headlen(skb), DMA_TO_DEVICE);
1397     txbuf->raw = 0;
1398 err_warn_dma:
1399     nn_dp_warn(dp, "Failed to map DMA TX buffer\n");
1400 err_free:
1401     u64_stats_update_begin(&r_vec->tx_sync);
1402     r_vec->tx_errors++;
1403     u64_stats_update_end(&r_vec->tx_sync);
1404     dev_kfree_skb_any(skb);
1405     return NETDEV_TX_OK;
1406 }
1407 
1408 static void __nfp_ctrl_tx_queued(struct nfp_net_r_vector *r_vec)
1409 {
1410     struct sk_buff *skb;
1411 
1412     while ((skb = __skb_dequeue(&r_vec->queue)))
1413         if (nfp_nfdk_ctrl_tx_one(r_vec->nfp_net, r_vec, skb, true))
1414             return;
1415 }
1416 
1417 static bool
1418 nfp_ctrl_meta_ok(struct nfp_net *nn, void *data, unsigned int meta_len)
1419 {
1420     u32 meta_type, meta_tag;
1421 
1422     if (!nfp_app_ctrl_has_meta(nn->app))
1423         return !meta_len;
1424 
1425     if (meta_len != 8)
1426         return false;
1427 
1428     meta_type = get_unaligned_be32(data);
1429     meta_tag = get_unaligned_be32(data + 4);
1430 
1431     return (meta_type == NFP_NET_META_PORTID &&
1432         meta_tag == NFP_META_PORT_ID_CTRL);
1433 }
1434 
1435 static bool
1436 nfp_ctrl_rx_one(struct nfp_net *nn, struct nfp_net_dp *dp,
1437         struct nfp_net_r_vector *r_vec, struct nfp_net_rx_ring *rx_ring)
1438 {
1439     unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
1440     struct nfp_net_rx_buf *rxbuf;
1441     struct nfp_net_rx_desc *rxd;
1442     dma_addr_t new_dma_addr;
1443     struct sk_buff *skb;
1444     void *new_frag;
1445     int idx;
1446 
1447     idx = D_IDX(rx_ring, rx_ring->rd_p);
1448 
1449     rxd = &rx_ring->rxds[idx];
1450     if (!(rxd->rxd.meta_len_dd & PCIE_DESC_RX_DD))
1451         return false;
1452 
1453     /* Memory barrier to ensure that we won't do other reads
1454      * before the DD bit.
1455      */
1456     dma_rmb();
1457 
1458     rx_ring->rd_p++;
1459 
1460     rxbuf = &rx_ring->rxbufs[idx];
1461     meta_len = rxd->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK;
1462     data_len = le16_to_cpu(rxd->rxd.data_len);
1463     pkt_len = data_len - meta_len;
1464 
1465     pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
1466     if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
1467         pkt_off += meta_len;
1468     else
1469         pkt_off += dp->rx_offset;
1470     meta_off = pkt_off - meta_len;
1471 
1472     /* Stats update */
1473     u64_stats_update_begin(&r_vec->rx_sync);
1474     r_vec->rx_pkts++;
1475     r_vec->rx_bytes += pkt_len;
1476     u64_stats_update_end(&r_vec->rx_sync);
1477 
1478     nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off, data_len);
1479 
1480     if (unlikely(!nfp_ctrl_meta_ok(nn, rxbuf->frag + meta_off, meta_len))) {
1481         nn_dp_warn(dp, "incorrect metadata for ctrl packet (%d)\n",
1482                meta_len);
1483         nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1484         return true;
1485     }
1486 
1487     skb = build_skb(rxbuf->frag, dp->fl_bufsz);
1488     if (unlikely(!skb)) {
1489         nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, NULL);
1490         return true;
1491     }
1492     new_frag = nfp_nfdk_napi_alloc_one(dp, &new_dma_addr);
1493     if (unlikely(!new_frag)) {
1494         nfp_nfdk_rx_drop(dp, r_vec, rx_ring, rxbuf, skb);
1495         return true;
1496     }
1497 
1498     nfp_net_dma_unmap_rx(dp, rxbuf->dma_addr);
1499 
1500     nfp_nfdk_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
1501 
1502     skb_reserve(skb, pkt_off);
1503     skb_put(skb, pkt_len);
1504 
1505     nfp_app_ctrl_rx(nn->app, skb);
1506 
1507     return true;
1508 }
1509 
1510 static bool nfp_ctrl_rx(struct nfp_net_r_vector *r_vec)
1511 {
1512     struct nfp_net_rx_ring *rx_ring = r_vec->rx_ring;
1513     struct nfp_net *nn = r_vec->nfp_net;
1514     struct nfp_net_dp *dp = &nn->dp;
1515     unsigned int budget = 512;
1516 
1517     while (nfp_ctrl_rx_one(nn, dp, r_vec, rx_ring) && budget--)
1518         continue;
1519 
1520     return budget;
1521 }
1522 
1523 void nfp_nfdk_ctrl_poll(struct tasklet_struct *t)
1524 {
1525     struct nfp_net_r_vector *r_vec = from_tasklet(r_vec, t, tasklet);
1526 
1527     spin_lock(&r_vec->lock);
1528     nfp_nfdk_tx_complete(r_vec->tx_ring, 0);
1529     __nfp_ctrl_tx_queued(r_vec);
1530     spin_unlock(&r_vec->lock);
1531 
1532     if (nfp_ctrl_rx(r_vec)) {
1533         nfp_net_irq_unmask(r_vec->nfp_net, r_vec->irq_entry);
1534     } else {
1535         tasklet_schedule(&r_vec->tasklet);
1536         nn_dp_warn(&r_vec->nfp_net->dp,
1537                "control message budget exceeded!\n");
1538     }
1539 }
[ 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!