| Message ID | 1408544322-32023-1-git-send-email-jonas.jensen@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, 2014-08-20 at 16:18 +0200, Jonas Jensen wrote: > TX buffer length is not cleared on ndo_start_xmit(). > Failing to do so can bug/hang the controller and > cause TX interrupts to stop altogether. > > diff --git a/drivers/net/ethernet/moxa/moxart_ether.c b/drivers/net/ethernet/moxa/moxart_ether.c > index 5020fd4..aa45607 100644 > --- a/drivers/net/ethernet/moxa/moxart_ether.c > +++ b/drivers/net/ethernet/moxa/moxart_ether.c > @@ -348,7 +348,8 @@ static int moxart_mac_start_xmit(struct sk_buff *skb, struct net_device *ndev) > > txdes1 = readl(desc + TX_REG_OFFSET_DESC1); > txdes1 |= TX_DESC1_LTS | TX_DESC1_FTS; > - txdes1 &= ~(TX_DESC1_FIFO_COMPLETE | TX_DESC1_INTR_COMPLETE); > + txdes1 &= ~(TX_DESC1_FIFO_COMPLETE | TX_DESC1_INTR_COMPLETE | > + TX_DESC1_BUF_SIZE_MASK); > txdes1 |= (len & TX_DESC1_BUF_SIZE_MASK); > writel(txdes1, desc + TX_REG_OFFSET_DESC1); > writel(TX_DESC0_DMA_OWN, desc + TX_REG_OFFSET_DESC0); Wouldnt it be faster to not use readl() at all here ? readl() is only used to get TX_DESC1_END bit. It seems you could replace the thing by : txdes1 = TX_DESC1_LTS | TX_DESC1_FTS | (len & TX_DESC1_BUF_SIZE_MASK); if (tx_head == TX_DESC_NUM_MASK) txdes1 |= TX_DESC1_END;
From: Jonas Jensen <jonas.jensen@gmail.com> Date: Wed, 20 Aug 2014 16:18:42 +0200 > @@ -348,7 +348,8 @@ static int moxart_mac_start_xmit(struct sk_buff *skb, struct net_device *ndev) > > txdes1 = readl(desc + TX_REG_OFFSET_DESC1); > txdes1 |= TX_DESC1_LTS | TX_DESC1_FTS; > - txdes1 &= ~(TX_DESC1_FIFO_COMPLETE | TX_DESC1_INTR_COMPLETE); > + txdes1 &= ~(TX_DESC1_FIFO_COMPLETE | TX_DESC1_INTR_COMPLETE | > + TX_DESC1_BUF_SIZE_MASK); > txdes1 |= (len & TX_DESC1_BUF_SIZE_MASK); > writel(txdes1, desc + TX_REG_OFFSET_DESC1); > writel(TX_DESC0_DMA_OWN, desc + TX_REG_OFFSET_DESC0); Like others I wonder why the existing descriptor value is being read at all. It's inefficient and completely unnecessary, you can just compute a new value from scratch, and that way all of these "uncleared field" issues just automatically disappear.
Thanks for giving feedback. On 20 August 2014 19:10, Eric Dumazet <eric.dumazet@gmail.com> wrote: > Wouldnt it be faster to not use readl() at all here ? On 22 August 2014 06:39, David Miller <davem@davemloft.net> wrote: > Like others I wonder why the existing descriptor value is being read > at all. It makes sense and I have verified, transfers continue to work clearing the whole thing, see update in v6. Regards, Jonas
diff --git a/drivers/net/ethernet/moxa/moxart_ether.c b/drivers/net/ethernet/moxa/moxart_ether.c index 5020fd4..aa45607 100644 --- a/drivers/net/ethernet/moxa/moxart_ether.c +++ b/drivers/net/ethernet/moxa/moxart_ether.c @@ -348,7 +348,8 @@ static int moxart_mac_start_xmit(struct sk_buff *skb, struct net_device *ndev) txdes1 = readl(desc + TX_REG_OFFSET_DESC1); txdes1 |= TX_DESC1_LTS | TX_DESC1_FTS; - txdes1 &= ~(TX_DESC1_FIFO_COMPLETE | TX_DESC1_INTR_COMPLETE); + txdes1 &= ~(TX_DESC1_FIFO_COMPLETE | TX_DESC1_INTR_COMPLETE | + TX_DESC1_BUF_SIZE_MASK); txdes1 |= (len & TX_DESC1_BUF_SIZE_MASK); writel(txdes1, desc + TX_REG_OFFSET_DESC1); writel(TX_DESC0_DMA_OWN, desc + TX_REG_OFFSET_DESC0);
TX buffer length is not cleared on ndo_start_xmit(). Failing to do so can bug/hang the controller and cause TX interrupts to stop altogether. Add TX_DESC1_BUF_SIZE_MASK to bits that are cleared, before the TX buffer length is set. Addresses https://bugzilla.kernel.org/show_bug.cgi?id=69031 Signed-off-by: Jonas Jensen <jonas.jensen@gmail.com> --- Notes: No changes since v4. Applies to next-20140820 drivers/net/ethernet/moxa/moxart_ether.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)