| Message ID | 1474023554-24520-1-git-send-email-vkuznets@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
From: Vitaly Kuznetsov <vkuznets@redhat.com> Date: Fri, 16 Sep 2016 12:59:14 +0200 > @@ -595,6 +596,19 @@ static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) > offset = offset_in_page(skb->data); > len = skb_headlen(skb); > > + /* The first req should be at least ETH_HLEN size or the packet will be > + * dropped by netback. > + */ > + if (unlikely(PAGE_SIZE - offset < ETH_HLEN)) { > + nskb = skb_copy(skb, GFP_ATOMIC); > + if (!nskb) > + goto drop; > + dev_kfree_skb_any(skb); > + skb = nskb; > + page = virt_to_page(skb->data); > + offset = offset_in_page(skb->data); > + } > + > spin_lock_irqsave(&queue->tx_lock, flags); I think you also have to recalculate 'len' in this case too, as skb_headlen() will definitely be different for nskb. In fact, I can't see how this code can work properly without that fix.
David Miller <davem@davemloft.net> writes: > From: Vitaly Kuznetsov <vkuznets@redhat.com> > Date: Fri, 16 Sep 2016 12:59:14 +0200 > >> @@ -595,6 +596,19 @@ static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) >> offset = offset_in_page(skb->data); >> len = skb_headlen(skb); >> >> + /* The first req should be at least ETH_HLEN size or the packet will be >> + * dropped by netback. >> + */ >> + if (unlikely(PAGE_SIZE - offset < ETH_HLEN)) { >> + nskb = skb_copy(skb, GFP_ATOMIC); >> + if (!nskb) >> + goto drop; >> + dev_kfree_skb_any(skb); >> + skb = nskb; >> + page = virt_to_page(skb->data); >> + offset = offset_in_page(skb->data); >> + } >> + >> spin_lock_irqsave(&queue->tx_lock, flags); > > I think you also have to recalculate 'len' in this case too, as > skb_headlen() will definitely be different for nskb. > > In fact, I can't see how this code can work properly without that fix. Thank you for your feedback David, in my testing (even when I tried doing skb_copy() for all skbs unconditionally) skb_headlen(nskb) always equals 'len' so I was under an impression that both 'skb->len' and 'skb->data_len' remain the same when we do skb_copy(). However, in case you think there are cases when headlen changes, I see no problem with re-calculating 'len' as it won't bring any significant performace penalty compared to the already added skb_copy(). I'll send 'v2'.
On 19/09/16 11:22, Vitaly Kuznetsov wrote: > David Miller <davem@davemloft.net> writes: > >> From: Vitaly Kuznetsov <vkuznets@redhat.com> >> Date: Fri, 16 Sep 2016 12:59:14 +0200 >> >>> @@ -595,6 +596,19 @@ static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) >>> offset = offset_in_page(skb->data); >>> len = skb_headlen(skb); >>> >>> + /* The first req should be at least ETH_HLEN size or the packet will be >>> + * dropped by netback. >>> + */ >>> + if (unlikely(PAGE_SIZE - offset < ETH_HLEN)) { >>> + nskb = skb_copy(skb, GFP_ATOMIC); >>> + if (!nskb) >>> + goto drop; >>> + dev_kfree_skb_any(skb); >>> + skb = nskb; >>> + page = virt_to_page(skb->data); >>> + offset = offset_in_page(skb->data); >>> + } >>> + >>> spin_lock_irqsave(&queue->tx_lock, flags); >> >> I think you also have to recalculate 'len' in this case too, as >> skb_headlen() will definitely be different for nskb. >> >> In fact, I can't see how this code can work properly without that fix. > > Thank you for your feedback David, > > in my testing (even when I tried doing skb_copy() for all skbs > unconditionally) skb_headlen(nskb) always equals 'len' so I was under an > impression that both 'skb->len' and 'skb->data_len' remain the same when > we do skb_copy(). However, in case you think there are cases when > headlen changes, I see no problem with re-calculating 'len' as it won't > bring any significant performace penalty compared to the already added > skb_copy(). I think you can move the len = skb_headlen(skb) after the if, no need to recalculate it. David
David Vrabel <david.vrabel@citrix.com> writes: > On 19/09/16 11:22, Vitaly Kuznetsov wrote: >> David Miller <davem@davemloft.net> writes: >> >>> From: Vitaly Kuznetsov <vkuznets@redhat.com> >>> Date: Fri, 16 Sep 2016 12:59:14 +0200 >>> >>>> @@ -595,6 +596,19 @@ static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) >>>> offset = offset_in_page(skb->data); >>>> len = skb_headlen(skb); >>>> >>>> + /* The first req should be at least ETH_HLEN size or the packet will be >>>> + * dropped by netback. >>>> + */ >>>> + if (unlikely(PAGE_SIZE - offset < ETH_HLEN)) { >>>> + nskb = skb_copy(skb, GFP_ATOMIC); >>>> + if (!nskb) >>>> + goto drop; >>>> + dev_kfree_skb_any(skb); >>>> + skb = nskb; >>>> + page = virt_to_page(skb->data); >>>> + offset = offset_in_page(skb->data); >>>> + } >>>> + >>>> spin_lock_irqsave(&queue->tx_lock, flags); >>> >>> I think you also have to recalculate 'len' in this case too, as >>> skb_headlen() will definitely be different for nskb. >>> >>> In fact, I can't see how this code can work properly without that fix. >> >> Thank you for your feedback David, >> >> in my testing (even when I tried doing skb_copy() for all skbs >> unconditionally) skb_headlen(nskb) always equals 'len' so I was under an >> impression that both 'skb->len' and 'skb->data_len' remain the same when >> we do skb_copy(). However, in case you think there are cases when >> headlen changes, I see no problem with re-calculating 'len' as it won't >> bring any significant performace penalty compared to the already added >> skb_copy(). > > I think you can move the len = skb_headlen(skb) after the if, no need to > recalculate it. Sorry, I was too fast with sending 'v2' and did the other way around. I'll do v3.
diff --git a/drivers/net/xen-netfront.c b/drivers/net/xen-netfront.c index 96ccd4e..28c4a66 100644 --- a/drivers/net/xen-netfront.c +++ b/drivers/net/xen-netfront.c @@ -565,6 +565,7 @@ static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) struct netfront_queue *queue = NULL; unsigned int num_queues = dev->real_num_tx_queues; u16 queue_index; + struct sk_buff *nskb; /* Drop the packet if no queues are set up */ if (num_queues < 1) @@ -595,6 +596,19 @@ static int xennet_start_xmit(struct sk_buff *skb, struct net_device *dev) offset = offset_in_page(skb->data); len = skb_headlen(skb); + /* The first req should be at least ETH_HLEN size or the packet will be + * dropped by netback. + */ + if (unlikely(PAGE_SIZE - offset < ETH_HLEN)) { + nskb = skb_copy(skb, GFP_ATOMIC); + if (!nskb) + goto drop; + dev_kfree_skb_any(skb); + skb = nskb; + page = virt_to_page(skb->data); + offset = offset_in_page(skb->data); + } + spin_lock_irqsave(&queue->tx_lock, flags); if (unlikely(!netif_carrier_ok(dev) ||
Small packet loss is reported on complex multi host network configurations including tunnels, NAT, ... My investigation led me to the following check in netback which drops packets: if (unlikely(txreq.size < ETH_HLEN)) { netdev_err(queue->vif->dev, "Bad packet size: %d\n", txreq.size); xenvif_tx_err(queue, &txreq, extra_count, idx); break; } But this check itself is legitimate. SKBs consist of a linear part (which has to have the ethernet header) and (optionally) a number of frags. Netfront transmits the head of the linear part up to the page boundary as the first request and all the rest becomes frags so when we're reconstructing the SKB in netback we can't distinguish between original frags and the 'tail' of the linear part. The first SKB needs to be at least ETH_HLEN size. So in case we have an SKB with its linear part starting too close to the page boundary the packet is lost. I see two ways to fix the issue: - Change the 'wire' protocol between netfront and netback to start keeping the original SKB structure. We'll have to add a flag indicating the fact that the particular request is a part of the original linear part and not a frag. We'll need to know the length of the linear part to pre-allocate memory. - Avoid transmitting SKBs with linear parts starting too close to the page boundary. That seems preferable short-term and shouldn't bring significant performance degradation as such packets are rare. That's what this patch is trying to achieve with skb_copy(). Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com> Acked-by: David Vrabel <david.vrabel@citrix.com> --- - This is just a RESEND with David's ACK added. --- drivers/net/xen-netfront.c | 14 ++++++++++++++ 1 file changed, 14 insertions(+)