dpaa2-eth: Use a single page per Rx buffer
authorIoana Ciocoi Radulescu <ruxandra.radulescu@nxp.com>
Mon, 4 Feb 2019 17:00:35 +0000 (17:00 +0000)
committerDavid S. Miller <davem@davemloft.net>
Wed, 6 Feb 2019 18:49:55 +0000 (10:49 -0800)
Instead of allocating page fragments via the network stack,
use the page allocator directly. For now, we consume one page
for each Rx buffer.

With the new memory model we are free to consider adding more
XDP support.

Performance decreases slightly in some IP forwarding cases.
No visible effect on termination traffic. The driver memory
footprint increases as a result of this change, but it is
still small enough to not really matter.

Another side effect is that now Rx buffer alignment requirements
are naturally satisfied without any additional actions needed.
Remove alignment related code, except in the buffer layout
information conveyed to MC, as hardware still needs to know the
alignment value we guarantee.

Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Signed-off-by: Ioana Radulescu <ruxandra.radulescu@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h

index 04925c731f0bb354afcdc1069ba93b83f2e9e7d6..6e58de628e5e713d5d5eed7c4628c50b8f5863c4 100644 (file)
@@ -86,16 +86,16 @@ static void free_rx_fd(struct dpaa2_eth_priv *priv,
        for (i = 1; i < DPAA2_ETH_MAX_SG_ENTRIES; i++) {
                addr = dpaa2_sg_get_addr(&sgt[i]);
                sg_vaddr = dpaa2_iova_to_virt(priv->iommu_domain, addr);
-               dma_unmap_single(dev, addr, DPAA2_ETH_RX_BUF_SIZE,
-                                DMA_BIDIRECTIONAL);
+               dma_unmap_page(dev, addr, DPAA2_ETH_RX_BUF_SIZE,
+                              DMA_BIDIRECTIONAL);
 
-               skb_free_frag(sg_vaddr);
+               free_pages((unsigned long)sg_vaddr, 0);
                if (dpaa2_sg_is_final(&sgt[i]))
                        break;
        }
 
 free_buf:
-       skb_free_frag(vaddr);
+       free_pages((unsigned long)vaddr, 0);
 }
 
 /* Build a linear skb based on a single-buffer frame descriptor */
@@ -109,7 +109,7 @@ static struct sk_buff *build_linear_skb(struct dpaa2_eth_channel *ch,
 
        ch->buf_count--;
 
-       skb = build_skb(fd_vaddr, DPAA2_ETH_SKB_SIZE);
+       skb = build_skb(fd_vaddr, DPAA2_ETH_RX_BUF_RAW_SIZE);
        if (unlikely(!skb))
                return NULL;
 
@@ -144,19 +144,19 @@ static struct sk_buff *build_frag_skb(struct dpaa2_eth_priv *priv,
                /* Get the address and length from the S/G entry */
                sg_addr = dpaa2_sg_get_addr(sge);
                sg_vaddr = dpaa2_iova_to_virt(priv->iommu_domain, sg_addr);
-               dma_unmap_single(dev, sg_addr, DPAA2_ETH_RX_BUF_SIZE,
-                                DMA_BIDIRECTIONAL);
+               dma_unmap_page(dev, sg_addr, DPAA2_ETH_RX_BUF_SIZE,
+                              DMA_BIDIRECTIONAL);
 
                sg_length = dpaa2_sg_get_len(sge);
 
                if (i == 0) {
                        /* We build the skb around the first data buffer */
-                       skb = build_skb(sg_vaddr, DPAA2_ETH_SKB_SIZE);
+                       skb = build_skb(sg_vaddr, DPAA2_ETH_RX_BUF_RAW_SIZE);
                        if (unlikely(!skb)) {
                                /* Free the first SG entry now, since we already
                                 * unmapped it and obtained the virtual address
                                 */
-                               skb_free_frag(sg_vaddr);
+                               free_pages((unsigned long)sg_vaddr, 0);
 
                                /* We still need to subtract the buffers used
                                 * by this FD from our software counter
@@ -211,9 +211,9 @@ static void free_bufs(struct dpaa2_eth_priv *priv, u64 *buf_array, int count)
 
        for (i = 0; i < count; i++) {
                vaddr = dpaa2_iova_to_virt(priv->iommu_domain, buf_array[i]);
-               dma_unmap_single(dev, buf_array[i], DPAA2_ETH_RX_BUF_SIZE,
-                                DMA_BIDIRECTIONAL);
-               skb_free_frag(vaddr);
+               dma_unmap_page(dev, buf_array[i], DPAA2_ETH_RX_BUF_SIZE,
+                              DMA_BIDIRECTIONAL);
+               free_pages((unsigned long)vaddr, 0);
        }
 }
 
@@ -378,16 +378,16 @@ static void dpaa2_eth_rx(struct dpaa2_eth_priv *priv,
                        return;
                }
 
-               dma_unmap_single(dev, addr, DPAA2_ETH_RX_BUF_SIZE,
-                                DMA_BIDIRECTIONAL);
+               dma_unmap_page(dev, addr, DPAA2_ETH_RX_BUF_SIZE,
+                              DMA_BIDIRECTIONAL);
                skb = build_linear_skb(ch, fd, vaddr);
        } else if (fd_format == dpaa2_fd_sg) {
                WARN_ON(priv->xdp_prog);
 
-               dma_unmap_single(dev, addr, DPAA2_ETH_RX_BUF_SIZE,
-                                DMA_BIDIRECTIONAL);
+               dma_unmap_page(dev, addr, DPAA2_ETH_RX_BUF_SIZE,
+                              DMA_BIDIRECTIONAL);
                skb = build_frag_skb(priv, ch, buf_data);
-               skb_free_frag(vaddr);
+               free_pages((unsigned long)vaddr, 0);
                percpu_extras->rx_sg_frames++;
                percpu_extras->rx_sg_bytes += dpaa2_fd_get_len(fd);
        } else {
@@ -903,7 +903,7 @@ static int add_bufs(struct dpaa2_eth_priv *priv,
 {
        struct device *dev = priv->net_dev->dev.parent;
        u64 buf_array[DPAA2_ETH_BUFS_PER_CMD];
-       void *buf;
+       struct page *page;
        dma_addr_t addr;
        int i, err;
 
@@ -911,14 +911,16 @@ static int add_bufs(struct dpaa2_eth_priv *priv,
                /* Allocate buffer visible to WRIOP + skb shared info +
                 * alignment padding
                 */
-               buf = napi_alloc_frag(dpaa2_eth_buf_raw_size(priv));
-               if (unlikely(!buf))
+               /* allocate one page for each Rx buffer. WRIOP sees
+                * the entire page except for a tailroom reserved for
+                * skb shared info
+                */
+               page = dev_alloc_pages(0);
+               if (!page)
                        goto err_alloc;
 
-               buf = PTR_ALIGN(buf, priv->rx_buf_align);
-
-               addr = dma_map_single(dev, buf, DPAA2_ETH_RX_BUF_SIZE,
-                                     DMA_BIDIRECTIONAL);
+               addr = dma_map_page(dev, page, 0, DPAA2_ETH_RX_BUF_SIZE,
+                                   DMA_BIDIRECTIONAL);
                if (unlikely(dma_mapping_error(dev, addr)))
                        goto err_map;
 
@@ -926,7 +928,7 @@ static int add_bufs(struct dpaa2_eth_priv *priv,
 
                /* tracing point */
                trace_dpaa2_eth_buf_seed(priv->net_dev,
-                                        buf, dpaa2_eth_buf_raw_size(priv),
+                                        page, DPAA2_ETH_RX_BUF_RAW_SIZE,
                                         addr, DPAA2_ETH_RX_BUF_SIZE,
                                         bpid);
        }
@@ -948,7 +950,7 @@ release_bufs:
        return i;
 
 err_map:
-       skb_free_frag(buf);
+       __free_pages(page, 0);
 err_alloc:
        /* If we managed to allocate at least some buffers,
         * release them to hardware
@@ -2134,6 +2136,7 @@ static int set_buffer_layout(struct dpaa2_eth_priv *priv)
 {
        struct device *dev = priv->net_dev->dev.parent;
        struct dpni_buffer_layout buf_layout = {0};
+       u16 rx_buf_align;
        int err;
 
        /* We need to check for WRIOP version 1.0.0, but depending on the MC
@@ -2142,9 +2145,9 @@ static int set_buffer_layout(struct dpaa2_eth_priv *priv)
         */
        if (priv->dpni_attrs.wriop_version == DPAA2_WRIOP_VERSION(0, 0, 0) ||
            priv->dpni_attrs.wriop_version == DPAA2_WRIOP_VERSION(1, 0, 0))
-               priv->rx_buf_align = DPAA2_ETH_RX_BUF_ALIGN_REV1;
+               rx_buf_align = DPAA2_ETH_RX_BUF_ALIGN_REV1;
        else
-               priv->rx_buf_align = DPAA2_ETH_RX_BUF_ALIGN;
+               rx_buf_align = DPAA2_ETH_RX_BUF_ALIGN;
 
        /* tx buffer */
        buf_layout.private_data_size = DPAA2_ETH_SWA_SIZE;
@@ -2184,7 +2187,7 @@ static int set_buffer_layout(struct dpaa2_eth_priv *priv)
        /* rx buffer */
        buf_layout.pass_frame_status = true;
        buf_layout.pass_parser_result = true;
-       buf_layout.data_align = priv->rx_buf_align;
+       buf_layout.data_align = rx_buf_align;
        buf_layout.data_head_room = dpaa2_eth_rx_head_room(priv);
        buf_layout.private_data_size = 0;
        buf_layout.options = DPNI_BUF_LAYOUT_OPT_PARSER_RESULT |
index 31fe486ec25f8df1fab4e346f6ca8f930f41ec52..da3d039cc9bfd0553e5e345fc50f2efef4607c67 100644 (file)
 /* Hardware requires alignment for ingress/egress buffer addresses */
 #define DPAA2_ETH_TX_BUF_ALIGN         64
 
-#define DPAA2_ETH_RX_BUF_SIZE          2048
-#define DPAA2_ETH_SKB_SIZE \
-       (DPAA2_ETH_RX_BUF_SIZE + SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
+#define DPAA2_ETH_RX_BUF_RAW_SIZE      PAGE_SIZE
+#define DPAA2_ETH_RX_BUF_TAILROOM \
+       SKB_DATA_ALIGN(sizeof(struct skb_shared_info))
+#define DPAA2_ETH_RX_BUF_SIZE \
+       (DPAA2_ETH_RX_BUF_RAW_SIZE - DPAA2_ETH_RX_BUF_TAILROOM)
 
 /* Hardware annotation area in RX/TX buffers */
 #define DPAA2_ETH_RX_HWA_SIZE          64
@@ -343,7 +345,6 @@ struct dpaa2_eth_priv {
        bool rx_tstamp; /* Rx timestamping enabled */
 
        u16 tx_qdid;
-       u16 rx_buf_align;
        struct fsl_mc_io *mc_io;
        /* Cores which have an affine DPIO/DPCON.
         * This is the cpu set on which Rx and Tx conf frames are processed
@@ -418,15 +419,6 @@ enum dpaa2_eth_rx_dist {
        DPAA2_ETH_RX_DIST_CLS
 };
 
-/* Hardware only sees DPAA2_ETH_RX_BUF_SIZE, but the skb built around
- * the buffer also needs space for its shared info struct, and we need
- * to allocate enough to accommodate hardware alignment restrictions
- */
-static inline unsigned int dpaa2_eth_buf_raw_size(struct dpaa2_eth_priv *priv)
-{
-       return DPAA2_ETH_SKB_SIZE + priv->rx_buf_align;
-}
-
 static inline
 unsigned int dpaa2_eth_needed_headroom(struct dpaa2_eth_priv *priv,
                                       struct sk_buff *skb)
@@ -451,8 +443,7 @@ unsigned int dpaa2_eth_needed_headroom(struct dpaa2_eth_priv *priv,
  */
 static inline unsigned int dpaa2_eth_rx_head_room(struct dpaa2_eth_priv *priv)
 {
-       return priv->tx_data_offset + DPAA2_ETH_TX_BUF_ALIGN -
-              DPAA2_ETH_RX_HWA_SIZE;
+       return priv->tx_data_offset - DPAA2_ETH_RX_HWA_SIZE;
 }
 
 int dpaa2_eth_set_hash(struct net_device *net_dev, u64 flags);