From mboxrd@z Thu Jan 1 00:00:00 1970 From: Sergei Shtylyov Date: Thu, 13 Nov 2014 23:05:44 +0000 Subject: Re: [PATCH 3/3] sh_eth: Fix dma mapping issue Message-Id: <54653948.5070606@cogentembedded.com> List-Id: References: <1415862301-28032-1-git-send-email-ykaneko0929@gmail.com> <1415862301-28032-4-git-send-email-ykaneko0929@gmail.com> In-Reply-To: <1415862301-28032-4-git-send-email-ykaneko0929@gmail.com> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Yoshihiro Kaneko , netdev@vger.kernel.org Cc: "David S. Miller" , Simon Horman , Magnus Damm , linux-sh@vger.kernel.org On 11/13/2014 10:05 AM, Yoshihiro Kaneko wrote: > From: Mitsuhiro Kimura > When CONFIG_DMA_API_DEBUG=y, many DMA error messages reports. > In order to use DMA debug, This patch fix following issues. > Issue 1: > If dma_mapping_error function is not called appropriately after > DMA mapping, DMA debug will report error message when DMA unmap > function is called. > Issue 2: > If skb_reserve function is called after DMA mapping, the relationship > between mapping addr and mapping size will be broken. > In this case, DMA debug will report error messages when DMA sync > function and DMA unmap function are called. > Issue 3: > If the size of frame data is less than ETH_ZLEN, the size is resized > to ETH_ZLEN after DMA map function is called. > In the TX skb freeing function, dma unmap function is called with that > resized value. So, unmap size error will reported. > Issue 4: > In the rx function, DMA map function is called without DMA unmap function > is called for RX skb reallocating. > It will case the DMA debug error that number of debug entry is full and > DMA debug logic is stopped. The rule of thumb is "fix one issue per patch". Please split accordingly. > Signed-off-by: Mitsuhiro Kimura > Signed-off-by: Yoshihiro Kaneko Thanks for beating me to it. Fixing these issues has been on my agenda for a long time... :-) > --- > drivers/net/ethernet/renesas/sh_eth.c | 26 +++++++++++++++++++++++--- > 1 file changed, 23 insertions(+), 3 deletions(-) > diff --git a/drivers/net/ethernet/renesas/sh_eth.c b/drivers/net/ethernet/renesas/sh_eth.c > index 0e4a407..23318cf 100644 > --- a/drivers/net/ethernet/renesas/sh_eth.c > +++ b/drivers/net/ethernet/renesas/sh_eth.c > @@ -1136,6 +1136,11 @@ static void sh_eth_ring_format(struct net_device *ndev) > dma_map_single(&ndev->dev, skb->data, rxdesc->buffer_length, > DMA_FROM_DEVICE); > rxdesc->addr = virt_to_phys(skb->data); Can't we get rid of these bogus virt_to_phys() calls, while at it? dma_map_single() returns a DMA address, no? > + if (dma_mapping_error(&ndev->dev, rxdesc->addr)) { > + dev_kfree_skb(mdp->rx_skbuff[i]); > + mdp->rx_skbuff[i] = NULL; > + break; > + } > rxdesc->status = cpu_to_edmac(mdp, RD_RACT | RD_RFP); > > /* Rx descriptor address set */ > @@ -1364,7 +1369,7 @@ static int sh_eth_txfree(struct net_device *ndev) > if (mdp->tx_skbuff[entry]) { > dma_unmap_single(&ndev->dev, txdesc->addr, > txdesc->buffer_length, DMA_TO_DEVICE); > - dev_kfree_skb_irq(mdp->tx_skbuff[entry]); > + dev_kfree_skb_any(mdp->tx_skbuff[entry]); Hm, I'm not sure where is this described in the changelog... > mdp->tx_skbuff[entry] = NULL; > free_num++; > } > @@ -1466,11 +1471,19 @@ static int sh_eth_rx(struct net_device *ndev, u32 intr_status, int *quota) > if (skb = NULL) > break; /* Better luck next round. */ > sh_eth_set_receive_align(skb); > + dma_unmap_single(&ndev->dev, rxdesc->addr, > + rxdesc->buffer_length, > + DMA_FROM_DEVICE); > dma_map_single(&ndev->dev, skb->data, > rxdesc->buffer_length, DMA_FROM_DEVICE); > > skb_checksum_none_assert(skb); > rxdesc->addr = virt_to_phys(skb->data); Likewise, can we get rid of this bogu? > + if (dma_mapping_error(&ndev->dev, rxdesc->addr)) { > + dev_kfree_skb_any(mdp->rx_skbuff[entry]); > + mdp->rx_skbuff[entry] = NULL; > + break; > + } > } > if (entry >= mdp->num_rx_ring - 1) > rxdesc->status |> @@ -2104,12 +2117,18 @@ static int sh_eth_start_xmit(struct sk_buff *skb, struct net_device *ndev) > if (!mdp->cd->hw_swap) > sh_eth_soft_swap(phys_to_virt(ALIGN(txdesc->addr, 4)), > skb->len + 2); > - txdesc->addr = dma_map_single(&ndev->dev, skb->data, skb->len, > - DMA_TO_DEVICE); > if (skb->len < ETH_ZLEN) > txdesc->buffer_length = ETH_ZLEN; > else > txdesc->buffer_length = skb->len; > + txdesc->addr = dma_map_single(&ndev->dev, skb->data, > + txdesc->buffer_length, > + DMA_TO_DEVICE); > + if (dma_mapping_error(&ndev->dev, txdesc->addr)) { > + dev_kfree_skb_any(mdp->tx_skbuff[entry]); > + mdp->tx_skbuff[entry] = NULL; > + goto out; Why not just *return*?! [...] WBR, Sergei From mboxrd@z Thu Jan 1 00:00:00 1970 From: Sergei Shtylyov Subject: Re: [PATCH 3/3] sh_eth: Fix dma mapping issue Date: Fri, 14 Nov 2014 02:05:44 +0300 Message-ID: <54653948.5070606@cogentembedded.com> References: <1415862301-28032-1-git-send-email-ykaneko0929@gmail.com> <1415862301-28032-4-git-send-email-ykaneko0929@gmail.com> Mime-Version: 1.0 Content-Type: text/plain; charset=windows-1252; format=flowed Content-Transfer-Encoding: 7bit Cc: "David S. Miller" , Simon Horman , Magnus Damm , linux-sh@vger.kernel.org To: Yoshihiro Kaneko , netdev@vger.kernel.org Return-path: In-Reply-To: <1415862301-28032-4-git-send-email-ykaneko0929@gmail.com> Sender: linux-sh-owner@vger.kernel.org List-Id: netdev.vger.kernel.org On 11/13/2014 10:05 AM, Yoshihiro Kaneko wrote: > From: Mitsuhiro Kimura > When CONFIG_DMA_API_DEBUG=y, many DMA error messages reports. > In order to use DMA debug, This patch fix following issues. > Issue 1: > If dma_mapping_error function is not called appropriately after > DMA mapping, DMA debug will report error message when DMA unmap > function is called. > Issue 2: > If skb_reserve function is called after DMA mapping, the relationship > between mapping addr and mapping size will be broken. > In this case, DMA debug will report error messages when DMA sync > function and DMA unmap function are called. > Issue 3: > If the size of frame data is less than ETH_ZLEN, the size is resized > to ETH_ZLEN after DMA map function is called. > In the TX skb freeing function, dma unmap function is called with that > resized value. So, unmap size error will reported. > Issue 4: > In the rx function, DMA map function is called without DMA unmap function > is called for RX skb reallocating. > It will case the DMA debug error that number of debug entry is full and > DMA debug logic is stopped. The rule of thumb is "fix one issue per patch". Please split accordingly. > Signed-off-by: Mitsuhiro Kimura > Signed-off-by: Yoshihiro Kaneko Thanks for beating me to it. Fixing these issues has been on my agenda for a long time... :-) > --- > drivers/net/ethernet/renesas/sh_eth.c | 26 +++++++++++++++++++++++--- > 1 file changed, 23 insertions(+), 3 deletions(-) > diff --git a/drivers/net/ethernet/renesas/sh_eth.c b/drivers/net/ethernet/renesas/sh_eth.c > index 0e4a407..23318cf 100644 > --- a/drivers/net/ethernet/renesas/sh_eth.c > +++ b/drivers/net/ethernet/renesas/sh_eth.c > @@ -1136,6 +1136,11 @@ static void sh_eth_ring_format(struct net_device *ndev) > dma_map_single(&ndev->dev, skb->data, rxdesc->buffer_length, > DMA_FROM_DEVICE); > rxdesc->addr = virt_to_phys(skb->data); Can't we get rid of these bogus virt_to_phys() calls, while at it? dma_map_single() returns a DMA address, no? > + if (dma_mapping_error(&ndev->dev, rxdesc->addr)) { > + dev_kfree_skb(mdp->rx_skbuff[i]); > + mdp->rx_skbuff[i] = NULL; > + break; > + } > rxdesc->status = cpu_to_edmac(mdp, RD_RACT | RD_RFP); > > /* Rx descriptor address set */ > @@ -1364,7 +1369,7 @@ static int sh_eth_txfree(struct net_device *ndev) > if (mdp->tx_skbuff[entry]) { > dma_unmap_single(&ndev->dev, txdesc->addr, > txdesc->buffer_length, DMA_TO_DEVICE); > - dev_kfree_skb_irq(mdp->tx_skbuff[entry]); > + dev_kfree_skb_any(mdp->tx_skbuff[entry]); Hm, I'm not sure where is this described in the changelog... > mdp->tx_skbuff[entry] = NULL; > free_num++; > } > @@ -1466,11 +1471,19 @@ static int sh_eth_rx(struct net_device *ndev, u32 intr_status, int *quota) > if (skb == NULL) > break; /* Better luck next round. */ > sh_eth_set_receive_align(skb); > + dma_unmap_single(&ndev->dev, rxdesc->addr, > + rxdesc->buffer_length, > + DMA_FROM_DEVICE); > dma_map_single(&ndev->dev, skb->data, > rxdesc->buffer_length, DMA_FROM_DEVICE); > > skb_checksum_none_assert(skb); > rxdesc->addr = virt_to_phys(skb->data); Likewise, can we get rid of this bogu? > + if (dma_mapping_error(&ndev->dev, rxdesc->addr)) { > + dev_kfree_skb_any(mdp->rx_skbuff[entry]); > + mdp->rx_skbuff[entry] = NULL; > + break; > + } > } > if (entry >= mdp->num_rx_ring - 1) > rxdesc->status |= > @@ -2104,12 +2117,18 @@ static int sh_eth_start_xmit(struct sk_buff *skb, struct net_device *ndev) > if (!mdp->cd->hw_swap) > sh_eth_soft_swap(phys_to_virt(ALIGN(txdesc->addr, 4)), > skb->len + 2); > - txdesc->addr = dma_map_single(&ndev->dev, skb->data, skb->len, > - DMA_TO_DEVICE); > if (skb->len < ETH_ZLEN) > txdesc->buffer_length = ETH_ZLEN; > else > txdesc->buffer_length = skb->len; > + txdesc->addr = dma_map_single(&ndev->dev, skb->data, > + txdesc->buffer_length, > + DMA_TO_DEVICE); > + if (dma_mapping_error(&ndev->dev, txdesc->addr)) { > + dev_kfree_skb_any(mdp->tx_skbuff[entry]); > + mdp->tx_skbuff[entry] = NULL; > + goto out; Why not just *return*?! [...] WBR, Sergei