]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/commitdiff
net: netsec: Use page_pool API
authorIlias Apalodimas <ilias.apalodimas@linaro.org>
Sat, 29 Jun 2019 05:23:23 +0000 (08:23 +0300)
committerDavid S. Miller <davem@davemloft.net>
Tue, 2 Jul 2019 02:27:08 +0000 (19:27 -0700)
Use page_pool and it's DMA mapping capabilities for Rx buffers instead
of netdev/napi_alloc_frag()

Although this will result in a slight performance penalty on small sized
packets (~10%) the use of the API will allow to easily add XDP support.
The penalty won't be visible in network testing i.e ipef/netperf etc, it
only happens during raw packet drops.
Furthermore we intend to add recycling capabilities on the API
in the future. Once the recycling is added the performance penalty will
go away.
The only 'real' penalty is the slightly increased memory usage, since we
now allocate a page per packet instead of the amount of bytes we need +
skb metadata (difference is roughly 2kb per packet).
With a minimum of 4BG of RAM on the only SoC that has this NIC the
extra memory usage is negligible (a bit more on 64K pages)

Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/socionext/Kconfig
drivers/net/ethernet/socionext/netsec.c

index 25f18be274239bf9bd54fbc9160dc80d6f45c066..95e99baf3f451f750136991a4383a9954fe75cc1 100644 (file)
@@ -26,6 +26,7 @@ config SNI_NETSEC
        tristate "Socionext NETSEC ethernet support"
        depends on (ARCH_SYNQUACER || COMPILE_TEST) && OF
        select PHYLIB
+       select PAGE_POOL
        select MII
        ---help---
          Enable to add support for the SocioNext NetSec Gigabit Ethernet
index 48fd7448b513fb9604e2b27b77e402f5d543c96e..7791bff2f2af3a6a5cbef5ec3b51dea29fe71c6b 100644 (file)
@@ -11,6 +11,7 @@
 #include <linux/io.h>
 
 #include <net/tcp.h>
+#include <net/page_pool.h>
 #include <net/ip6_checksum.h>
 
 #define NETSEC_REG_SOFT_RST                    0x104
 #define DESC_NUM       256
 
 #define NETSEC_SKB_PAD (NET_SKB_PAD + NET_IP_ALIGN)
-#define NETSEC_RX_BUF_SZ 1536
+#define NETSEC_RX_BUF_NON_DATA (NETSEC_SKB_PAD + \
+                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
 
 #define DESC_SZ        sizeof(struct netsec_de)
 
@@ -258,6 +260,8 @@ struct netsec_desc_ring {
        struct netsec_desc *desc;
        void *vaddr;
        u16 head, tail;
+       struct page_pool *page_pool;
+       struct xdp_rxq_info xdp_rxq;
 };
 
 struct netsec_priv {
@@ -673,33 +677,27 @@ static void netsec_process_tx(struct netsec_priv *priv)
 }
 
 static void *netsec_alloc_rx_data(struct netsec_priv *priv,
-                                 dma_addr_t *dma_handle, u16 *desc_len,
-                                 bool napi)
+                                 dma_addr_t *dma_handle, u16 *desc_len)
+
 {
-       size_t total_len = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-       size_t payload_len = NETSEC_RX_BUF_SZ;
-       dma_addr_t mapping;
-       void *buf;
 
-       total_len += SKB_DATA_ALIGN(payload_len + NETSEC_SKB_PAD);
+       struct netsec_desc_ring *dring = &priv->desc_ring[NETSEC_RING_RX];
+       struct page *page;
 
-       buf = napi ? napi_alloc_frag(total_len) : netdev_alloc_frag(total_len);
-       if (!buf)
+       page = page_pool_dev_alloc_pages(dring->page_pool);
+       if (!page)
                return NULL;
 
-       mapping = dma_map_single(priv->dev, buf + NETSEC_SKB_PAD, payload_len,
-                                DMA_FROM_DEVICE);
-       if (unlikely(dma_mapping_error(priv->dev, mapping)))
-               goto err_out;
-
-       *dma_handle = mapping;
-       *desc_len = payload_len;
-
-       return buf;
+       /* page_pool API will map the whole page, skip
+        * NET_SKB_PAD + NET_IP_ALIGN for the payload
+        */
+       *dma_handle = page_pool_get_dma_addr(page) + NETSEC_SKB_PAD;
+       /* make sure the incoming payload fits in the page with the needed
+        * NET_SKB_PAD + NET_IP_ALIGN + skb_shared_info
+        */
+       *desc_len = PAGE_SIZE - NETSEC_RX_BUF_NON_DATA;
 
-err_out:
-       skb_free_frag(buf);
-       return NULL;
+       return page_address(page);
 }
 
 static void netsec_rx_fill(struct netsec_priv *priv, u16 from, u16 num)
@@ -728,10 +726,10 @@ static int netsec_process_rx(struct netsec_priv *priv, int budget)
                u16 idx = dring->tail;
                struct netsec_de *de = dring->vaddr + (DESC_SZ * idx);
                struct netsec_desc *desc = &dring->desc[idx];
+               struct page *page = virt_to_page(desc->addr);
                u16 pkt_len, desc_len;
                dma_addr_t dma_handle;
                void *buf_addr;
-               u32 truesize;
 
                if (de->attr & (1U << NETSEC_RX_PKT_OWN_FIELD)) {
                        /* reading the register clears the irq */
@@ -766,8 +764,8 @@ static int netsec_process_rx(struct netsec_priv *priv, int budget)
                /* allocate a fresh buffer and map it to the hardware.
                 * This will eventually replace the old buffer in the hardware
                 */
-               buf_addr = netsec_alloc_rx_data(priv, &dma_handle, &desc_len,
-                                               true);
+               buf_addr = netsec_alloc_rx_data(priv, &dma_handle, &desc_len);
+
                if (unlikely(!buf_addr))
                        break;
 
@@ -775,22 +773,19 @@ static int netsec_process_rx(struct netsec_priv *priv, int budget)
                                        DMA_FROM_DEVICE);
                prefetch(desc->addr);
 
-               truesize = SKB_DATA_ALIGN(desc->len + NETSEC_SKB_PAD) +
-                       SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
-               skb = build_skb(desc->addr, truesize);
+               skb = build_skb(desc->addr, desc->len + NETSEC_RX_BUF_NON_DATA);
                if (unlikely(!skb)) {
-                       /* free the newly allocated buffer, we are not going to
-                        * use it
+                       /* If skb fails recycle_direct will either unmap and
+                        * free the page or refill the cache depending on the
+                        * cache state. Since we paid the allocation cost if
+                        * building an skb fails try to put the page into cache
                         */
-                       dma_unmap_single(priv->dev, dma_handle, desc_len,
-                                        DMA_FROM_DEVICE);
-                       skb_free_frag(buf_addr);
+                       page_pool_recycle_direct(dring->page_pool, page);
                        netif_err(priv, drv, priv->ndev,
                                  "rx failed to build skb\n");
                        break;
                }
-               dma_unmap_single_attrs(priv->dev, desc->dma_addr, desc->len,
-                                      DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
+               page_pool_release_page(dring->page_pool, page);
 
                /* Update the descriptor with the new buffer we allocated */
                desc->len = desc_len;
@@ -980,19 +975,31 @@ static void netsec_uninit_pkt_dring(struct netsec_priv *priv, int id)
 
        if (!dring->vaddr || !dring->desc)
                return;
-
        for (idx = 0; idx < DESC_NUM; idx++) {
                desc = &dring->desc[idx];
                if (!desc->addr)
                        continue;
 
-               dma_unmap_single(priv->dev, desc->dma_addr, desc->len,
-                                id == NETSEC_RING_RX ? DMA_FROM_DEVICE :
-                                                             DMA_TO_DEVICE);
-               if (id == NETSEC_RING_RX)
-                       skb_free_frag(desc->addr);
-               else if (id == NETSEC_RING_TX)
+               if (id == NETSEC_RING_RX) {
+                       struct page *page = virt_to_page(desc->addr);
+
+                       page_pool_put_page(dring->page_pool, page, false);
+               } else if (id == NETSEC_RING_TX) {
+                       dma_unmap_single(priv->dev, desc->dma_addr, desc->len,
+                                        DMA_TO_DEVICE);
                        dev_kfree_skb(desc->skb);
+               }
+       }
+
+       /* Rx is currently using page_pool
+        * since the pool is created during netsec_setup_rx_dring(), we need to
+        * free the pool manually if the registration failed
+        */
+       if (id == NETSEC_RING_RX) {
+               if (xdp_rxq_info_is_reg(&dring->xdp_rxq))
+                       xdp_rxq_info_unreg(&dring->xdp_rxq);
+               else
+                       page_pool_free(dring->page_pool);
        }
 
        memset(dring->desc, 0, sizeof(struct netsec_desc) * DESC_NUM);
@@ -1059,7 +1066,23 @@ static void netsec_setup_tx_dring(struct netsec_priv *priv)
 static int netsec_setup_rx_dring(struct netsec_priv *priv)
 {
        struct netsec_desc_ring *dring = &priv->desc_ring[NETSEC_RING_RX];
-       int i;
+       struct page_pool_params pp_params = { 0 };
+       int i, err;
+
+       pp_params.order = 0;
+       /* internal DMA mapping in page_pool */
+       pp_params.flags = PP_FLAG_DMA_MAP;
+       pp_params.pool_size = DESC_NUM;
+       pp_params.nid = cpu_to_node(0);
+       pp_params.dev = priv->dev;
+       pp_params.dma_dir = DMA_FROM_DEVICE;
+
+       dring->page_pool = page_pool_create(&pp_params);
+       if (IS_ERR(dring->page_pool)) {
+               err = PTR_ERR(dring->page_pool);
+               dring->page_pool = NULL;
+               goto err_out;
+       }
 
        for (i = 0; i < DESC_NUM; i++) {
                struct netsec_desc *desc = &dring->desc[i];
@@ -1067,10 +1090,10 @@ static int netsec_setup_rx_dring(struct netsec_priv *priv)
                void *buf;
                u16 len;
 
-               buf = netsec_alloc_rx_data(priv, &dma_handle, &len,
-                                          false);
+               buf = netsec_alloc_rx_data(priv, &dma_handle, &len);
+
                if (!buf) {
-                       netsec_uninit_pkt_dring(priv, NETSEC_RING_RX);
+                       err = -ENOMEM;
                        goto err_out;
                }
                desc->dma_addr = dma_handle;
@@ -1079,11 +1102,20 @@ static int netsec_setup_rx_dring(struct netsec_priv *priv)
        }
 
        netsec_rx_fill(priv, 0, DESC_NUM);
+       err = xdp_rxq_info_reg(&dring->xdp_rxq, priv->ndev, 0);
+       if (err)
+               goto err_out;
+
+       err = xdp_rxq_info_reg_mem_model(&dring->xdp_rxq, MEM_TYPE_PAGE_POOL,
+                                        dring->page_pool);
+       if (err)
+               goto err_out;
 
        return 0;
 
 err_out:
-       return -ENOMEM;
+       netsec_uninit_pkt_dring(priv, NETSEC_RING_RX);
+       return err;
 }
 
 static int netsec_netdev_load_ucode_region(struct netsec_priv *priv, u32 reg,