Message ID | 1425086169.5130.57.camel@edumazet-glaptop2.roam.corp.google.com |
---|---|
State | RFC, archived |
Delegated to: | David Miller |
Headers | show |
On 2/27/2015 8:16 PM, Eric Dumazet wrote: > On Fri, 2015-02-27 at 16:48 -0800, Eric Dumazet wrote: >> On Fri, 2015-02-27 at 16:37 -0500, Brian Rak wrote: >>> I've been seeing a crash under 3.19.0 that seems to occur when I put >>> heavy traffic across a macvtap/veth interface. >>> >>> We have a KVM guest attached to a veth pair using macvtap. We're >>> routing IPv6 traffic into one end of the veth pair using some static >>> routes. We do *not* have proxy_ndp enabled (though, we are using some >>> software to do neighbor proxying - http://priv.nu/projects/ndppd/ ). >>> >>> I've been able to reproduce this pretty easily by downloading some large >>> files from the guest. We see two traces in a row when this occurs: >> >> >> Nice ! >> >> Crash is in neigh_hh_output() >> >> -> memcpy(skb->data - HH_DATA_MOD, hh->hh_data, HH_DATA_MOD); >> >> And there is only 14 bytes of headroom instead of 16. >> >> Some layer did not align skb_headroom(skb) to HH_DATA_MOD for ethernet >> header. > > Could you try following patch ? > > diff --git a/drivers/net/macvtap.c b/drivers/net/macvtap.c > index e40fdfccc9c10df4ea8676a1dd59275d5d9c6b88..27ecc5c4fa2665cd42ac1ca81717255f85507113 100644 > --- a/drivers/net/macvtap.c > +++ b/drivers/net/macvtap.c > @@ -654,11 +654,14 @@ static void macvtap_skb_to_vnet_hdr(struct macvtap_queue *q, > } /* else everything is zero */ > } > > +/* Neighbour code has some assumptions on HH_DATA_MOD alignment */ > +#define MACVTAP_RESERVE HH_DATA_OFF(ETH_HLEN) > + > /* Get packet from user space buffer */ > static ssize_t macvtap_get_user(struct macvtap_queue *q, struct msghdr *m, > struct iov_iter *from, int noblock) > { > - int good_linear = SKB_MAX_HEAD(NET_IP_ALIGN); > + int good_linear = SKB_MAX_HEAD(MACVTAP_RESERVE); > struct sk_buff *skb; > struct macvlan_dev *vlan; > unsigned long total_len = iov_iter_count(from); > @@ -722,7 +725,7 @@ static ssize_t macvtap_get_user(struct macvtap_queue *q, struct msghdr *m, > linear = macvtap16_to_cpu(q, vnet_hdr.hdr_len); > } > > - skb = macvtap_alloc_skb(&q->sk, NET_IP_ALIGN, copylen, > + skb = macvtap_alloc_skb(&q->sk, MACVTAP_RESERVE, copylen, > linear, noblock, &err); > if (!skb) > goto err; > > Wow, that was *much* faster then I was expecting, thanks a bunch! I can confirm that resolves the issue.. I've tested this and it fixes the issue perfectly. I've been able to put a whole bunch of IPv6 traffic through the interface now, whereas before even a minor amount of traffic would crash the host. Thanks again! -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 2015-02-27 at 20:54 -0500, Brian Rak wrote: > Wow, that was *much* faster then I was expecting, thanks a bunch! > > I can confirm that resolves the issue.. I've tested this and it fixes > the issue perfectly. I've been able to put a whole bunch of IPv6 > traffic through the interface now, whereas before even a minor amount of > traffic would crash the host. > > Thanks again! Interesting... Had a prior version of linux kernel been fine ? -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 2015-02-27 at 18:01 -0800, Eric Dumazet wrote: > On Fri, 2015-02-27 at 20:54 -0500, Brian Rak wrote: > > > Wow, that was *much* faster then I was expecting, thanks a bunch! > > > > I can confirm that resolves the issue.. I've tested this and it fixes > > the issue perfectly. I've been able to put a whole bunch of IPv6 > > traffic through the interface now, whereas before even a minor amount of > > traffic would crash the host. > > > > Thanks again! > > Interesting... > > Had a prior version of linux kernel been fine ? Or maybe you recently switched on this config option ? CONFIG_DEBUG_PAGEALLOC=y -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 2/27/2015 9:03 PM, Eric Dumazet wrote: > On Fri, 2015-02-27 at 18:01 -0800, Eric Dumazet wrote: >> On Fri, 2015-02-27 at 20:54 -0500, Brian Rak wrote: >> >>> Wow, that was *much* faster then I was expecting, thanks a bunch! >>> >>> I can confirm that resolves the issue.. I've tested this and it fixes >>> the issue perfectly. I've been able to put a whole bunch of IPv6 >>> traffic through the interface now, whereas before even a minor amount of >>> traffic would crash the host. >>> >>> Thanks again! >> >> Interesting... >> >> Had a prior version of linux kernel been fine ? > > Or maybe you recently switched on this config option ? > > CONFIG_DEBUG_PAGEALLOC=y > > > We've only recently started using this veth/macvtap combo, so it's possible this has been around for awhile and we just hadn't noticed. I don't have any info on older kernels currently. I *think* I've seen crashes on 3.17.1, but I didn't save any stack traces, so I can't be sure. CONFIG_DEBUG_PAGEALLOC is not set, and never has been. -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, 2015-02-27 at 21:11 -0500, Brian Rak wrote: > We've only recently started using this veth/macvtap combo, so it's > possible this has been around for awhile and we just hadn't noticed. > > I don't have any info on older kernels currently. I *think* I've seen > crashes on 3.17.1, but I didn't save any stack traces, so I can't be sure. > > CONFIG_DEBUG_PAGEALLOC is not set, and never has been. OK, thanks for the confirmation. I'll send an official patch. (I guess same patch is also needed for drivers/net/tun.c) -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/net/macvtap.c b/drivers/net/macvtap.c index e40fdfccc9c10df4ea8676a1dd59275d5d9c6b88..27ecc5c4fa2665cd42ac1ca81717255f85507113 100644 --- a/drivers/net/macvtap.c +++ b/drivers/net/macvtap.c @@ -654,11 +654,14 @@ static void macvtap_skb_to_vnet_hdr(struct macvtap_queue *q, } /* else everything is zero */ } +/* Neighbour code has some assumptions on HH_DATA_MOD alignment */ +#define MACVTAP_RESERVE HH_DATA_OFF(ETH_HLEN) + /* Get packet from user space buffer */ static ssize_t macvtap_get_user(struct macvtap_queue *q, struct msghdr *m, struct iov_iter *from, int noblock) { - int good_linear = SKB_MAX_HEAD(NET_IP_ALIGN); + int good_linear = SKB_MAX_HEAD(MACVTAP_RESERVE); struct sk_buff *skb; struct macvlan_dev *vlan; unsigned long total_len = iov_iter_count(from); @@ -722,7 +725,7 @@ static ssize_t macvtap_get_user(struct macvtap_queue *q, struct msghdr *m, linear = macvtap16_to_cpu(q, vnet_hdr.hdr_len); } - skb = macvtap_alloc_skb(&q->sk, NET_IP_ALIGN, copylen, + skb = macvtap_alloc_skb(&q->sk, MACVTAP_RESERVE, copylen, linear, noblock, &err); if (!skb) goto err;