nfp: fix issue of skb segments exceeds descriptor limitation

TCP packets will be dropped if the segments number in the tx skb
exceeds limitation when sending iperf3 traffic with --zerocopy option.

we make the following changes:

Get nr_frags in nfp_nfdk_tx_maybe_close_block instead of passing from
outside because it will be changed after skb_linearize operation.

Fill maximum dma_len in first tx descriptor to make sure the whole
head is included in the first descriptor.

Fixes: 2c359c6702e5 ("nfp: add support for NFDK data path")
Signed-off-by: Baowen Zheng <baowen.zheng@corigine.com>
Reviewed-by: Louis Peens <louis.peens@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

diff --git a/drivers/net/ethernet/netronome/nfp/nfdk/dp.c b/drivers/net/ethernet/netronome/nfp/nfdk/dp.c
index e509d6dcba5cb3defee06d29f8157adf134deda8..805071d64a20388de19f5a49784e7c82178c6e36 100644 (file)
--- a/drivers/net/ethernet/netronome/nfp/nfdk/dp.c
+++ b/drivers/net/ethernet/netronome/nfp/nfdk/dp.c
@@ -125,17 +125,18 @@ nfp_nfdk_tx_csum(struct nfp_net_dp *dp, struct nfp_net_r_vector *r_vec,
 
 static int
 nfp_nfdk_tx_maybe_close_block(struct nfp_net_tx_ring *tx_ring,
-                             unsigned int nr_frags, struct sk_buff *skb)
+                             struct sk_buff *skb)
 {
        unsigned int n_descs, wr_p, nop_slots;
        const skb_frag_t *frag, *fend;
        struct nfp_nfdk_tx_desc *txd;
+       unsigned int nr_frags;
        unsigned int wr_idx;
        int err;
 
 recount_descs:
        n_descs = nfp_nfdk_headlen_to_segs(skb_headlen(skb));
-
+       nr_frags = skb_shinfo(skb)->nr_frags;
        frag = skb_shinfo(skb)->frags;
        fend = frag + nr_frags;
        for (; frag < fend; frag++)
@@ -281,10 +282,13 @@ netdev_tx_t nfp_nfdk_tx(struct sk_buff *skb, struct net_device *netdev)
        if (unlikely((int)metadata < 0))
                goto err_flush;
 
-       nr_frags = skb_shinfo(skb)->nr_frags;
-       if (nfp_nfdk_tx_maybe_close_block(tx_ring, nr_frags, skb))
+       if (nfp_nfdk_tx_maybe_close_block(tx_ring, skb))
                goto err_flush;
 
+       /* nr_frags will change after skb_linearize so we get nr_frags after
+        * nfp_nfdk_tx_maybe_close_block function
+        */
+       nr_frags = skb_shinfo(skb)->nr_frags;
        /* DMA map all */
        wr_idx = D_IDX(tx_ring, tx_ring->wr_p);
        txd = &tx_ring->ktxds[wr_idx];
@@ -310,7 +314,16 @@ netdev_tx_t nfp_nfdk_tx(struct sk_buff *skb, struct net_device *netdev)
 
        /* FIELD_PREP() implicitly truncates to chunk */
        dma_len -= 1;
-       dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD, dma_len) |
+
+       /* We will do our best to pass as much data as we can in descriptor
+        * and we need to make sure the first descriptor includes whole head
+        * since there is limitation in firmware side. Sometimes the value of
+        * dma_len bitwise and NFDK_DESC_TX_DMA_LEN_HEAD will less than
+        * headlen.
+        */
+       dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
+                              dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
+                              NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
                    FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
 
        txd->dma_len_type = cpu_to_le16(dlen_type);
@@ -925,7 +938,9 @@ nfp_nfdk_tx_xdp_buf(struct nfp_net_dp *dp, struct nfp_net_rx_ring *rx_ring,
 
        /* FIELD_PREP() implicitly truncates to chunk */
        dma_len -= 1;
-       dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD, dma_len) |
+       dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
+                              dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
+                              NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
                    FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
 
        txd->dma_len_type = cpu_to_le16(dlen_type);
@@ -1303,7 +1318,7 @@ nfp_nfdk_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
                                   skb_push(skb, 4));
        }
 
-       if (nfp_nfdk_tx_maybe_close_block(tx_ring, 0, skb))
+       if (nfp_nfdk_tx_maybe_close_block(tx_ring, skb))
                goto err_free;
 
        /* DMA map all */
@@ -1328,7 +1343,9 @@ nfp_nfdk_ctrl_tx_one(struct nfp_net *nn, struct nfp_net_r_vector *r_vec,
        txbuf++;
 
        dma_len -= 1;
-       dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD, dma_len) |
+       dlen_type = FIELD_PREP(NFDK_DESC_TX_DMA_LEN_HEAD,
+                              dma_len > NFDK_DESC_TX_DMA_LEN_HEAD ?
+                              NFDK_DESC_TX_DMA_LEN_HEAD : dma_len) |
                    FIELD_PREP(NFDK_DESC_TX_TYPE_HEAD, type);
 
        txd->dma_len_type = cpu_to_le16(dlen_type);