| Message ID | 1473697270-3906793-1-git-send-email-tom@herbertland.com |
|---|---|
| State | RFC, archived |
| Delegated to: | David Miller |
| Headers | show |
On Mon, 2016-09-12 at 09:21 -0700, Tom Herbert wrote: > In order to use XDP with non multi-queue drivers (such as e1000) we need > a method to handle XDP_TX return code from xdp program. Since there is > only one queue the XDP transmit path needs to share that queue with the > normal stack path, and in order to do that we can't break the existing > stack's mechanisms of locking, qdiscs, and BQL for the queue. > > This patch allocates an skbuff when XDP_TX is returned from the XDP > program and call dev_queue_xmit to transmit the skbuff. > > Note that this is expected to be a performance hit (which needs to be > quantified) relative to a MQ XDP implementation where we can reserve > queues to be used by XDP. The first intent of these patches is > correctness. This should not affect performance in the XDP_DROP path. > > This patch needs to be applied after "e1000: add initial XDP support". > > Tested: I've only built this, need to see if I can find some e1000 > machines. John, if you could try this that would be much appreciated. > > Signed-off-by: Tom Herbert <tom@herbertland.com> > --- > drivers/net/ethernet/intel/e1000/e1000_main.c | 29 ++++++++++++++++++++------- > 1 file changed, 22 insertions(+), 7 deletions(-) > > diff --git a/drivers/net/ethernet/intel/e1000/e1000_main.c b/drivers/net/ethernet/intel/e1000/e1000_main.c > index 91d5c87..f457ea8 100644 > --- a/drivers/net/ethernet/intel/e1000/e1000_main.c > +++ b/drivers/net/ethernet/intel/e1000/e1000_main.c > @@ -4338,13 +4338,28 @@ static bool e1000_clean_jumbo_rx_irq(struct e1000_adapter *adapter, > case XDP_PASS: > break; > case XDP_TX: > - dma_sync_single_for_device(&pdev->dev, > - dma, > - length, > - DMA_TO_DEVICE); > - e1000_xmit_raw_frame(buffer_info, length, > - netdev, adapter); > - buffer_info->rxbuf.page = NULL; > + /* There is only one transmit queue so we need > + * to inject the packet at the qdisc layer > + * to maintain sanity in the stack. We allocate > + * an skbuff, set up frags with the page, and > + * then do dev_queue_xmit. > + */ > + skb = alloc_skb(0, GFP_ATOMIC); > + if (skb) { > + dma_unmap_page(&pdev->dev, > + buffer_info->dma, > + adapter->rx_buffer_len, > + DMA_FROM_DEVICE); > + skb->dev = netdev; > + skb_fill_page_desc(skb, 0, > + buffer_info->rxbuf.page, > + 0, length); > + dev_queue_xmit(skb); > + goto next_desc; > + } > + /* Fallthrough, if we are unable to handle > + * XDP_TX then we need to drop. > + */ > case XDP_DROP: > default: > /* re-use mapped page. keep buffer_info->dma If the intent is correctness, then you'll need to populate skb->len and skb->data_len. Also many drivers assume some headers are in skb->head. e1000 seems to assume skb_headlen() can not be 0
On Mon, Sep 12, 2016 at 09:21:10AM -0700, Tom Herbert wrote: > In order to use XDP with non multi-queue drivers (such as e1000) we need > a method to handle XDP_TX return code from xdp program. Since there is > only one queue the XDP transmit path needs to share that queue with the > normal stack path, and in order to do that we can't break the existing > stack's mechanisms of locking, qdiscs, and BQL for the queue. > > This patch allocates an skbuff when XDP_TX is returned from the XDP > program and call dev_queue_xmit to transmit the skbuff. > > Note that this is expected to be a performance hit (which needs to be > quantified) relative to a MQ XDP implementation where we can reserve > queues to be used by XDP. The first intent of these patches is > correctness. This should not affect performance in the XDP_DROP path. > > This patch needs to be applied after "e1000: add initial XDP support". > > Tested: I've only built this, need to see if I can find some e1000 > machines. John, if you could try this that would be much appreciated. > > Signed-off-by: Tom Herbert <tom@herbertland.com> > --- > drivers/net/ethernet/intel/e1000/e1000_main.c | 29 ++++++++++++++++++++------- > 1 file changed, 22 insertions(+), 7 deletions(-) > > diff --git a/drivers/net/ethernet/intel/e1000/e1000_main.c b/drivers/net/ethernet/intel/e1000/e1000_main.c > index 91d5c87..f457ea8 100644 > --- a/drivers/net/ethernet/intel/e1000/e1000_main.c > +++ b/drivers/net/ethernet/intel/e1000/e1000_main.c > @@ -4338,13 +4338,28 @@ static bool e1000_clean_jumbo_rx_irq(struct e1000_adapter *adapter, > case XDP_PASS: > break; > case XDP_TX: > - dma_sync_single_for_device(&pdev->dev, > - dma, > - length, > - DMA_TO_DEVICE); > - e1000_xmit_raw_frame(buffer_info, length, > - netdev, adapter); > - buffer_info->rxbuf.page = NULL; > + /* There is only one transmit queue so we need > + * to inject the packet at the qdisc layer > + * to maintain sanity in the stack. We allocate > + * an skbuff, set up frags with the page, and > + * then do dev_queue_xmit. > + */ > + skb = alloc_skb(0, GFP_ATOMIC); > + if (skb) { > + dma_unmap_page(&pdev->dev, > + buffer_info->dma, > + adapter->rx_buffer_len, > + DMA_FROM_DEVICE); > + skb->dev = netdev; > + skb_fill_page_desc(skb, 0, > + buffer_info->rxbuf.page, > + 0, length); > + dev_queue_xmit(skb); I don't like this approach, since it does exactly what it meant to: It uses qdisc and everything else in xmit path of the stack. It's a serious issue, since for some drivers XDP_TX would mean raw xmit whereas for e1k it goes via stack. Sooner or later program authors will start relying on that and that's not acceptable from XDP design point of view.
diff --git a/drivers/net/ethernet/intel/e1000/e1000_main.c b/drivers/net/ethernet/intel/e1000/e1000_main.c index 91d5c87..f457ea8 100644 --- a/drivers/net/ethernet/intel/e1000/e1000_main.c +++ b/drivers/net/ethernet/intel/e1000/e1000_main.c @@ -4338,13 +4338,28 @@ static bool e1000_clean_jumbo_rx_irq(struct e1000_adapter *adapter, case XDP_PASS: break; case XDP_TX: - dma_sync_single_for_device(&pdev->dev, - dma, - length, - DMA_TO_DEVICE); - e1000_xmit_raw_frame(buffer_info, length, - netdev, adapter); - buffer_info->rxbuf.page = NULL; + /* There is only one transmit queue so we need + * to inject the packet at the qdisc layer + * to maintain sanity in the stack. We allocate + * an skbuff, set up frags with the page, and + * then do dev_queue_xmit. + */ + skb = alloc_skb(0, GFP_ATOMIC); + if (skb) { + dma_unmap_page(&pdev->dev, + buffer_info->dma, + adapter->rx_buffer_len, + DMA_FROM_DEVICE); + skb->dev = netdev; + skb_fill_page_desc(skb, 0, + buffer_info->rxbuf.page, + 0, length); + dev_queue_xmit(skb); + goto next_desc; + } + /* Fallthrough, if we are unable to handle + * XDP_TX then we need to drop. + */ case XDP_DROP: default: /* re-use mapped page. keep buffer_info->dma
In order to use XDP with non multi-queue drivers (such as e1000) we need a method to handle XDP_TX return code from xdp program. Since there is only one queue the XDP transmit path needs to share that queue with the normal stack path, and in order to do that we can't break the existing stack's mechanisms of locking, qdiscs, and BQL for the queue. This patch allocates an skbuff when XDP_TX is returned from the XDP program and call dev_queue_xmit to transmit the skbuff. Note that this is expected to be a performance hit (which needs to be quantified) relative to a MQ XDP implementation where we can reserve queues to be used by XDP. The first intent of these patches is correctness. This should not affect performance in the XDP_DROP path. This patch needs to be applied after "e1000: add initial XDP support". Tested: I've only built this, need to see if I can find some e1000 machines. John, if you could try this that would be much appreciated. Signed-off-by: Tom Herbert <tom@herbertland.com> --- drivers/net/ethernet/intel/e1000/e1000_main.c | 29 ++++++++++++++++++++------- 1 file changed, 22 insertions(+), 7 deletions(-)