net:fec: fix WARNING caused by lack of calls to dma_mapping_error()
authorDuan Fugang-B38611 <B38611@freescale.com>
Thu, 14 Nov 2013 01:57:10 +0000 (09:57 +0800)
committerDavid S. Miller <davem@davemloft.net>
Thu, 14 Nov 2013 21:37:33 +0000 (16:37 -0500)
The driver fails to check the results of DMA mapping and results in
the following warning: (with kernel config "CONFIG_DMA_API_DEBUG" enable)

------------[ cut here ]------------
WARNING: at lib/dma-debug.c:937 check_unmap+0x43c/0x7d8()
fec 2188000.ethernet: DMA-API: device driver failed to check map
error[device address=0x00000000383a8040] [size=2048 bytes] [mapped as single]

Modules linked in:
CPU: 0 PID: 0 Comm: swapper/0 Not tainted 3.10.17-16827-g9cdb0ba-dirty #188
[<80013c4c>] (unwind_backtrace+0x0/0xf8) from [<80011704>] (show_stack+0x10/0x14)
[<80011704>] (show_stack+0x10/0x14) from [<80025614>] (warn_slowpath_common+0x4c/0x6c)
[<80025614>] (warn_slowpath_common+0x4c/0x6c) from [<800256c8>] (warn_slowpath_fmt+0x30/0x40)
[<800256c8>] (warn_slowpath_fmt+0x30/0x40) from [<8026bfdc>] (check_unmap+0x43c/0x7d8)
[<8026bfdc>] (check_unmap+0x43c/0x7d8) from [<8026c584>] (debug_dma_unmap_page+0x6c/0x78)
[<8026c584>] (debug_dma_unmap_page+0x6c/0x78) from [<8038049c>] (fec_enet_rx_napi+0x254/0x8a8)
[<8038049c>] (fec_enet_rx_napi+0x254/0x8a8) from [<804dc8c0>] (net_rx_action+0x94/0x160)
[<804dc8c0>] (net_rx_action+0x94/0x160) from [<8002c758>] (__do_softirq+0xe8/0x1d0)
[<8002c758>] (__do_softirq+0xe8/0x1d0) from [<8002c8e8>] (do_softirq+0x4c/0x58)
[<8002c8e8>] (do_softirq+0x4c/0x58) from [<8002cb50>] (irq_exit+0x90/0xc8)
[<8002cb50>] (irq_exit+0x90/0xc8) from [<8000ea88>] (handle_IRQ+0x3c/0x94)
[<8000ea88>] (handle_IRQ+0x3c/0x94) from [<8000855c>] (gic_handle_irq+0x28/0x5c)
[<8000855c>] (gic_handle_irq+0x28/0x5c) from [<8000de00>] (__irq_svc+0x40/0x50)
Exception stack(0x815a5f38 to 0x815a5f80)
5f20:                                                       815a5f80 3b9aca00
5f40: 0fe52383 00000002 0dd8950e 00000002 81e7b080 00000000 00000000 815ac4d8
5f60: 806032ec 00000000 00000017 815a5f80 80059028 8041fc4c 60000013 ffffffff
[<8000de00>] (__irq_svc+0x40/0x50) from [<8041fc4c>] (cpuidle_enter_state+0x50/0xf0)
[<8041fc4c>] (cpuidle_enter_state+0x50/0xf0) from [<8041fd94>] (cpuidle_idle_call+0xa8/0x14c)
[<8041fd94>] (cpuidle_idle_call+0xa8/0x14c) from [<8000edac>] (arch_cpu_idle+0x10/0x4c)
[<8000edac>] (arch_cpu_idle+0x10/0x4c) from [<800582f8>] (cpu_startup_entry+0x60/0x130)
[<800582f8>] (cpu_startup_entry+0x60/0x130) from [<80bc7a48>] (start_kernel+0x2d0/0x328)
[<80bc7a48>] (start_kernel+0x2d0/0x328) from [<10008074>] (0x10008074)
---[ end trace c6edec32436e0042 ]---

Because dma-debug add new interfaces to debug dma mapping errors, pls refer
to: http://lwn.net/Articles/516640/

After dma mapping, it must call dma_mapping_error() to check mapping error,
otherwise the map_err_type alway is MAP_ERR_NOT_CHECKED, check_unmap() define
the mapping is not checked and dump the error msg. So,add dma_mapping_error()
checking to fix the WARNING

And RX DMA buffers are used repeatedly and the driver copies it into an skb,
fec_enet_rx() should not map or unmap, use dma_sync_single_for_cpu()/dma_sync_single_for_device()
instead of dma_map_single()/dma_unmap_single().

There have another potential issue:  fec_enet_rx() passes the DMA address to __va().
Physical and DMA addresses are *not* the same thing. They may differ if the device
is behind an IOMMU or bounce buffering was required, or just because there is a fixed
offset between the device and host physical addresses. Also fix it in this patch.

=============================================
V2: add net_ratelimit() to limit map err message.
    use dma_sync_single_for_cpu() instead of dma_map_single().
    fix the issue that pass DMA addresses to __va() to get virture address.
V1: initial send
=============================================

Signed-off-by: Fugang Duan <B38611@freescale.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/freescale/fec_main.c

index b2793b9..4cbebf3 100644 (file)
@@ -386,7 +386,14 @@ fec_enet_start_xmit(struct sk_buff *skb, struct net_device *ndev)
         */
        bdp->cbd_bufaddr = dma_map_single(&fep->pdev->dev, bufaddr,
                        FEC_ENET_TX_FRSIZE, DMA_TO_DEVICE);
-
+       if (dma_mapping_error(&fep->pdev->dev, bdp->cbd_bufaddr)) {
+               bdp->cbd_bufaddr = 0;
+               fep->tx_skbuff[index] = NULL;
+               dev_kfree_skb_any(skb);
+               if (net_ratelimit())
+                       netdev_err(ndev, "Tx DMA memory map failed\n");
+               return NETDEV_TX_OK;
+       }
        /* Send it on its way.  Tell FEC it's ready, interrupt when done,
         * it's the last BD of the frame, and to put the CRC on the end.
         */
@@ -861,6 +868,7 @@ fec_enet_rx(struct net_device *ndev, int budget)
        struct  bufdesc_ex *ebdp = NULL;
        bool    vlan_packet_rcvd = false;
        u16     vlan_tag;
+       int     index = 0;
 
 #ifdef CONFIG_M532x
        flush_cache_all();
@@ -916,10 +924,15 @@ fec_enet_rx(struct net_device *ndev, int budget)
                ndev->stats.rx_packets++;
                pkt_len = bdp->cbd_datlen;
                ndev->stats.rx_bytes += pkt_len;
-               data = (__u8*)__va(bdp->cbd_bufaddr);
 
-               dma_unmap_single(&fep->pdev->dev, bdp->cbd_bufaddr,
-                               FEC_ENET_TX_FRSIZE, DMA_FROM_DEVICE);
+               if (fep->bufdesc_ex)
+                       index = (struct bufdesc_ex *)bdp -
+                               (struct bufdesc_ex *)fep->rx_bd_base;
+               else
+                       index = bdp - fep->rx_bd_base;
+               data = fep->rx_skbuff[index]->data;
+               dma_sync_single_for_cpu(&fep->pdev->dev, bdp->cbd_bufaddr,
+                                       FEC_ENET_RX_FRSIZE, DMA_FROM_DEVICE);
 
                if (id_entry->driver_data & FEC_QUIRK_SWAP_FRAME)
                        swap_buffer(data, pkt_len);
@@ -999,8 +1012,8 @@ fec_enet_rx(struct net_device *ndev, int budget)
                        napi_gro_receive(&fep->napi, skb);
                }
 
-               bdp->cbd_bufaddr = dma_map_single(&fep->pdev->dev, data,
-                               FEC_ENET_TX_FRSIZE, DMA_FROM_DEVICE);
+               dma_sync_single_for_device(&fep->pdev->dev, bdp->cbd_bufaddr,
+                                       FEC_ENET_RX_FRSIZE, DMA_FROM_DEVICE);
 rx_processing_done:
                /* Clear the status flags for this buffer */
                status &= ~BD_ENET_RX_STATS;
@@ -1719,6 +1732,12 @@ static int fec_enet_alloc_buffers(struct net_device *ndev)
 
                bdp->cbd_bufaddr = dma_map_single(&fep->pdev->dev, skb->data,
                                FEC_ENET_RX_FRSIZE, DMA_FROM_DEVICE);
+               if (dma_mapping_error(&fep->pdev->dev, bdp->cbd_bufaddr)) {
+                       fec_enet_free_buffers(ndev);
+                       if (net_ratelimit())
+                               netdev_err(ndev, "Rx DMA memory map failed\n");
+                       return -ENOMEM;
+               }
                bdp->cbd_sc = BD_ENET_RX_EMPTY;
 
                if (fep->bufdesc_ex) {