nfp: make use of the DMA_ATTR_SKIP_CPU_SYNC attr
authorJakub Kicinski <jakub.kicinski@netronome.com>
Sun, 23 Apr 2017 03:17:52 +0000 (20:17 -0700)
committerDavid S. Miller <davem@davemloft.net>
Mon, 24 Apr 2017 18:35:44 +0000 (14:35 -0400)
DMA unmap may destroy changes CPU made to the buffer.  To make XDP
run correctly on non-x86 platforms we should use the
DMA_ATTR_SKIP_CPU_SYNC attribute.

Thanks to using the attribute we can now push the sync operation to the
common code path from XDP handler.

A little bit of variable name reshuffling is required to bring the
code back to readable state.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/netronome/nfp/nfp_net_common.c

index e2197160e4dcc70fe626f6f06c4d20d70925ebc7..f1128d12cd2462dcfe22b22091fb08d6bff71501 100644 (file)
@@ -87,16 +87,31 @@ void nfp_net_get_fw_version(struct nfp_net_fw_version *fw_ver,
 
 static dma_addr_t nfp_net_dma_map_rx(struct nfp_net_dp *dp, void *frag)
 {
-       return dma_map_single(dp->dev, frag + NFP_NET_RX_BUF_HEADROOM,
-                             dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
-                             dp->rx_dma_dir);
+       return dma_map_single_attrs(dp->dev, frag + NFP_NET_RX_BUF_HEADROOM,
+                                   dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
+                                   dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
+}
+
+static void
+nfp_net_dma_sync_dev_rx(const struct nfp_net_dp *dp, dma_addr_t dma_addr)
+{
+       dma_sync_single_for_device(dp->dev, dma_addr,
+                                  dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
+                                  dp->rx_dma_dir);
 }
 
 static void nfp_net_dma_unmap_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr)
 {
-       dma_unmap_single(dp->dev, dma_addr,
-                        dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
-                        dp->rx_dma_dir);
+       dma_unmap_single_attrs(dp->dev, dma_addr,
+                              dp->fl_bufsz - NFP_NET_RX_BUF_NON_DATA,
+                              dp->rx_dma_dir, DMA_ATTR_SKIP_CPU_SYNC);
+}
+
+static void nfp_net_dma_sync_cpu_rx(struct nfp_net_dp *dp, dma_addr_t dma_addr,
+                                   unsigned int len)
+{
+       dma_sync_single_for_cpu(dp->dev, dma_addr - NFP_NET_RX_BUF_HEADROOM,
+                               len, dp->rx_dma_dir);
 }
 
 /* Firmware reconfig
@@ -1208,6 +1223,8 @@ static void nfp_net_rx_give_one(const struct nfp_net_dp *dp,
 
        wr_idx = rx_ring->wr_p & (rx_ring->cnt - 1);
 
+       nfp_net_dma_sync_dev_rx(dp, dma_addr);
+
        /* Stash SKB and DMA address away */
        rx_ring->rxbufs[wr_idx].frag = frag;
        rx_ring->rxbufs[wr_idx].dma_addr = dma_addr;
@@ -1569,7 +1586,7 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
        tx_ring = r_vec->xdp_ring;
 
        while (pkts_polled < budget) {
-               unsigned int meta_len, data_len, data_off, pkt_len;
+               unsigned int meta_len, data_len, meta_off, pkt_len, pkt_off;
                u8 meta_prepend[NFP_NET_MAX_PREPEND];
                struct nfp_net_rx_buf *rxbuf;
                struct nfp_net_rx_desc *rxd;
@@ -1608,11 +1625,12 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
                data_len = le16_to_cpu(rxd->rxd.data_len);
                pkt_len = data_len - meta_len;
 
+               pkt_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_dma_off;
                if (dp->rx_offset == NFP_NET_CFG_RX_OFFSET_DYNAMIC)
-                       data_off = NFP_NET_RX_BUF_HEADROOM + meta_len;
+                       pkt_off += meta_len;
                else
-                       data_off = NFP_NET_RX_BUF_HEADROOM + dp->rx_offset;
-               data_off += dp->rx_dma_off;
+                       pkt_off += dp->rx_offset;
+               meta_off = pkt_off - meta_len;
 
                /* Stats update */
                u64_stats_update_begin(&r_vec->rx_sync);
@@ -1621,7 +1639,7 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
                u64_stats_update_end(&r_vec->rx_sync);
 
                /* Pointer to start of metadata */
-               meta = rxbuf->frag + data_off - meta_len;
+               meta = rxbuf->frag + meta_off;
 
                if (unlikely(meta_len > NFP_NET_MAX_PREPEND ||
                             (dp->rx_offset && meta_len > dp->rx_offset))) {
@@ -1631,6 +1649,9 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
                        continue;
                }
 
+               nfp_net_dma_sync_cpu_rx(dp, rxbuf->dma_addr + meta_off,
+                                       data_len);
+
                if (xdp_prog && !(rxd->rxd.flags & PCIE_DESC_RX_BPF &&
                                  dp->bpf_offload_xdp)) {
                        unsigned int dma_off;
@@ -1638,10 +1659,6 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
                        int act;
 
                        hard_start = rxbuf->frag + NFP_NET_RX_BUF_HEADROOM;
-                       dma_off = data_off - NFP_NET_RX_BUF_HEADROOM;
-                       dma_sync_single_for_cpu(dp->dev, rxbuf->dma_addr,
-                                               dma_off + pkt_len,
-                                               DMA_BIDIRECTIONAL);
 
                        /* Move prepend out of the way */
                        if (xdp_prog->xdp_adjust_head) {
@@ -1650,12 +1667,12 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
                        }
 
                        act = nfp_net_run_xdp(xdp_prog, rxbuf->frag, hard_start,
-                                             &data_off, &pkt_len);
+                                             &pkt_off, &pkt_len);
                        switch (act) {
                        case XDP_PASS:
                                break;
                        case XDP_TX:
-                               dma_off = data_off - NFP_NET_RX_BUF_HEADROOM;
+                               dma_off = pkt_off - NFP_NET_RX_BUF_HEADROOM;
                                if (unlikely(!nfp_net_tx_xdp_buf(dp, rx_ring,
                                                                 tx_ring, rxbuf,
                                                                 dma_off,
@@ -1689,7 +1706,7 @@ static int nfp_net_rx(struct nfp_net_rx_ring *rx_ring, int budget)
 
                nfp_net_rx_give_one(dp, rx_ring, new_frag, new_dma_addr);
 
-               skb_reserve(skb, data_off);
+               skb_reserve(skb, pkt_off);
                skb_put(skb, pkt_len);
 
                if (!dp->chained_metadata_format) {