net: mvneta: introduce page pool API for sw buffer manager
authorLorenzo Bianconi <lorenzo@kernel.org>
Sat, 19 Oct 2019 08:13:22 +0000 (10:13 +0200)
committerDavid S. Miller <davem@davemloft.net>
Mon, 21 Oct 2019 17:36:07 +0000 (10:36 -0700)
Use the page_pool api for allocations and DMA handling instead of
__dev_alloc_page()/dma_map_page() and free_page()/dma_unmap_page().
Pages are unmapped using page_pool_release_page before packets
go into the network stack.

The page_pool API offers buffer recycling capabilities for XDP but
allocates one page per packet, unless the driver splits and manages
the allocated page.
This is a preliminary patch to add XDP support to mvneta driver

Signed-off-by: Ilias Apalodimas <ilias.apalodimas@linaro.org>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/marvell/Kconfig
drivers/net/ethernet/marvell/mvneta.c

index fb942167ee54014ea1b4ed7b3a83594a278dfdfa..3d5caea096fbaa78c66d7b425056788ce6137356 100644 (file)
@@ -61,6 +61,7 @@ config MVNETA
        depends on ARCH_MVEBU || COMPILE_TEST
        select MVMDIO
        select PHYLINK
+       select PAGE_POOL
        ---help---
          This driver supports the network interface units in the
          Marvell ARMADA XP, ARMADA 370, ARMADA 38x and
index 3a8110d0a5b64f54bc4d707a96729127e51f5ae0..cd1a59c7e9d207cf1bad4e4a4f778aee683da0bb 100644 (file)
@@ -37,6 +37,7 @@
 #include <net/ip.h>
 #include <net/ipv6.h>
 #include <net/tso.h>
+#include <net/page_pool.h>
 
 /* Registers */
 #define MVNETA_RXQ_CONFIG_REG(q)                (0x1400 + ((q) << 2))
@@ -603,6 +604,10 @@ struct mvneta_rx_queue {
        u32 pkts_coal;
        u32 time_coal;
 
+       /* page_pool */
+       struct page_pool *page_pool;
+       struct xdp_rxq_info xdp_rxq;
+
        /* Virtual address of the RX buffer */
        void  **buf_virt_addr;
 
@@ -1812,23 +1817,21 @@ static int mvneta_rx_refill(struct mvneta_port *pp,
                            struct mvneta_rx_queue *rxq,
                            gfp_t gfp_mask)
 {
+       enum dma_data_direction dma_dir;
        dma_addr_t phys_addr;
        struct page *page;
 
-       page = __dev_alloc_page(gfp_mask);
+       page = page_pool_alloc_pages(rxq->page_pool,
+                                    gfp_mask | __GFP_NOWARN);
        if (!page)
                return -ENOMEM;
 
-       /* map page for use */
-       phys_addr = dma_map_page(pp->dev->dev.parent, page, 0, PAGE_SIZE,
-                                DMA_FROM_DEVICE);
-       if (unlikely(dma_mapping_error(pp->dev->dev.parent, phys_addr))) {
-               __free_page(page);
-               return -ENOMEM;
-       }
-
-       phys_addr += pp->rx_offset_correction;
+       phys_addr = page_pool_get_dma_addr(page) + pp->rx_offset_correction;
+       dma_dir = page_pool_get_dma_dir(rxq->page_pool);
+       dma_sync_single_for_device(pp->dev->dev.parent, phys_addr,
+                                  PAGE_SIZE, dma_dir);
        mvneta_rx_desc_fill(rx_desc, phys_addr, page, rxq);
+
        return 0;
 }
 
@@ -1894,10 +1897,12 @@ static void mvneta_rxq_drop_pkts(struct mvneta_port *pp,
                if (!data || !(rx_desc->buf_phys_addr))
                        continue;
 
-               dma_unmap_page(pp->dev->dev.parent, rx_desc->buf_phys_addr,
-                              PAGE_SIZE, DMA_FROM_DEVICE);
-               __free_page(data);
+               page_pool_put_page(rxq->page_pool, data, false);
        }
+       if (xdp_rxq_info_is_reg(&rxq->xdp_rxq))
+               xdp_rxq_info_unreg(&rxq->xdp_rxq);
+       page_pool_destroy(rxq->page_pool);
+       rxq->page_pool = NULL;
 }
 
 static void
@@ -2029,8 +2034,7 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
                                skb_add_rx_frag(rxq->skb, frag_num, page,
                                                frag_offset, frag_size,
                                                PAGE_SIZE);
-                               dma_unmap_page(dev->dev.parent, phys_addr,
-                                              PAGE_SIZE, DMA_FROM_DEVICE);
+                               page_pool_release_page(rxq->page_pool, page);
                                rxq->left_size -= frag_size;
                        }
                } else {
@@ -2060,9 +2064,7 @@ static int mvneta_rx_swbm(struct napi_struct *napi,
                                                frag_offset, frag_size,
                                                PAGE_SIZE);
 
-                               dma_unmap_page(dev->dev.parent, phys_addr,
-                                              PAGE_SIZE, DMA_FROM_DEVICE);
-
+                               page_pool_release_page(rxq->page_pool, page);
                                rxq->left_size -= frag_size;
                        }
                } /* Middle or Last descriptor */
@@ -2831,11 +2833,54 @@ static int mvneta_poll(struct napi_struct *napi, int budget)
        return rx_done;
 }
 
+static int mvneta_create_page_pool(struct mvneta_port *pp,
+                                  struct mvneta_rx_queue *rxq, int size)
+{
+       struct page_pool_params pp_params = {
+               .order = 0,
+               .flags = PP_FLAG_DMA_MAP,
+               .pool_size = size,
+               .nid = cpu_to_node(0),
+               .dev = pp->dev->dev.parent,
+               .dma_dir = DMA_FROM_DEVICE,
+       };
+       int err;
+
+       rxq->page_pool = page_pool_create(&pp_params);
+       if (IS_ERR(rxq->page_pool)) {
+               err = PTR_ERR(rxq->page_pool);
+               rxq->page_pool = NULL;
+               return err;
+       }
+
+       err = xdp_rxq_info_reg(&rxq->xdp_rxq, pp->dev, rxq->id);
+       if (err < 0)
+               goto err_free_pp;
+
+       err = xdp_rxq_info_reg_mem_model(&rxq->xdp_rxq, MEM_TYPE_PAGE_POOL,
+                                        rxq->page_pool);
+       if (err)
+               goto err_unregister_rxq;
+
+       return 0;
+
+err_unregister_rxq:
+       xdp_rxq_info_unreg(&rxq->xdp_rxq);
+err_free_pp:
+       page_pool_destroy(rxq->page_pool);
+       rxq->page_pool = NULL;
+       return err;
+}
+
 /* Handle rxq fill: allocates rxq skbs; called when initializing a port */
 static int mvneta_rxq_fill(struct mvneta_port *pp, struct mvneta_rx_queue *rxq,
                           int num)
 {
-       int i;
+       int i, err;
+
+       err = mvneta_create_page_pool(pp, rxq, num);
+       if (err < 0)
+               return err;
 
        for (i = 0; i < num; i++) {
                memset(rxq->descs + i, 0, sizeof(struct mvneta_rx_desc));