bpf: add skb_postpush_rcsum and fix dev_forward_skb occasions
authorDaniel Borkmann <daniel@iogearbox.net>
Thu, 7 Jan 2016 14:50:23 +0000 (15:50 +0100)
committerDavid S. Miller <davem@davemloft.net>
Sun, 10 Jan 2016 22:54:28 +0000 (17:54 -0500)
Add a small helper skb_postpush_rcsum() and fix up redirect locations
that need CHECKSUM_COMPLETE fixups on ingress. dev_forward_skb() expects
a proper csum that covers also Ethernet header, f.e. since 2c26d34bbcc0
("net/core: Handle csum for CHECKSUM_COMPLETE VXLAN forwarding"), we
also do skb_postpull_rcsum() after pulling Ethernet header off via
eth_type_trans().

When using eBPF in a netns setup f.e. with vxlan in collect metadata mode,
I can trigger the following csum issue with an IPv6 setup:

  [  505.144065] dummy1: hw csum failure
  [...]
  [  505.144108] Call Trace:
  [  505.144112]  <IRQ>  [<ffffffff81372f08>] dump_stack+0x44/0x5c
  [  505.144134]  [<ffffffff81607cea>] netdev_rx_csum_fault+0x3a/0x40
  [  505.144142]  [<ffffffff815fee3f>] __skb_checksum_complete+0xcf/0xe0
  [  505.144149]  [<ffffffff816f0902>] nf_ip6_checksum+0xb2/0x120
  [  505.144161]  [<ffffffffa08c0e0e>] icmpv6_error+0x17e/0x328 [nf_conntrack_ipv6]
  [  505.144170]  [<ffffffffa0898eca>] ? ip6t_do_table+0x2fa/0x645 [ip6_tables]
  [  505.144177]  [<ffffffffa08c0725>] ? ipv6_get_l4proto+0x65/0xd0 [nf_conntrack_ipv6]
  [  505.144189]  [<ffffffffa06c9a12>] nf_conntrack_in+0xc2/0x5a0 [nf_conntrack]
  [  505.144196]  [<ffffffffa08c039c>] ipv6_conntrack_in+0x1c/0x20 [nf_conntrack_ipv6]
  [  505.144204]  [<ffffffff8164385d>] nf_iterate+0x5d/0x70
  [  505.144210]  [<ffffffff816438d6>] nf_hook_slow+0x66/0xc0
  [  505.144218]  [<ffffffff816bd302>] ipv6_rcv+0x3f2/0x4f0
  [  505.144225]  [<ffffffff816bca40>] ? ip6_make_skb+0x1b0/0x1b0
  [  505.144232]  [<ffffffff8160b77b>] __netif_receive_skb_core+0x36b/0x9a0
  [  505.144239]  [<ffffffff8160bdc8>] ? __netif_receive_skb+0x18/0x60
  [  505.144245]  [<ffffffff8160bdc8>] __netif_receive_skb+0x18/0x60
  [  505.144252]  [<ffffffff8160ccff>] process_backlog+0x9f/0x140
  [  505.144259]  [<ffffffff8160c4a5>] net_rx_action+0x145/0x320
  [...]

What happens is that on ingress, we push Ethernet header back in, either
from cls_bpf or right before skb_do_redirect(), but without updating csum.
The "hw csum failure" can be fixed by using the new skb_postpush_rcsum()
helper for the dev_forward_skb() case to correct the csum diff again.

Thanks to Hannes Frederic Sowa for the csum_partial() idea!

Fixes: 3896d655f4d4 ("bpf: introduce bpf_clone_redirect() helper")
Fixes: 27b29f63058d ("bpf: add bpf_redirect() helper")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/skbuff.h
net/core/filter.c

index 6b6bd42d6134a3bfccb36f0bad4447c8c35ab3ff..07f9ccd28654b6590edb63b97b0aa42b1d4748cf 100644 (file)
@@ -2805,6 +2805,23 @@ static inline void skb_postpull_rcsum(struct sk_buff *skb,
 
 unsigned char *skb_pull_rcsum(struct sk_buff *skb, unsigned int len);
 
+static inline void skb_postpush_rcsum(struct sk_buff *skb,
+                                     const void *start, unsigned int len)
+{
+       /* For performing the reverse operation to skb_postpull_rcsum(),
+        * we can instead of ...
+        *
+        *   skb->csum = csum_add(skb->csum, csum_partial(start, len, 0));
+        *
+        * ... just use this equivalent version here to save a few
+        * instructions. Feeding csum of 0 in csum_partial() and later
+        * on adding skb->csum is equivalent to feed skb->csum in the
+        * first place.
+        */
+       if (skb->ip_summed == CHECKSUM_COMPLETE)
+               skb->csum = csum_partial(start, len, skb->csum);
+}
+
 /**
  *     pskb_trim_rcsum - trim received skb and update checksum
  *     @skb: buffer to trim
index 35e6fed287093b7f81eabf1b159a1efd542b6a29..0db92b5e2cbfb49364196586c88dbec7aa741523 100644 (file)
@@ -1368,8 +1368,9 @@ static u64 bpf_skb_store_bytes(u64 r1, u64 r2, u64 r3, u64 r4, u64 flags)
                /* skb_store_bits cannot return -EFAULT here */
                skb_store_bits(skb, offset, ptr, len);
 
-       if (BPF_RECOMPUTE_CSUM(flags) && skb->ip_summed == CHECKSUM_COMPLETE)
-               skb->csum = csum_add(skb->csum, csum_partial(ptr, len, 0));
+       if (BPF_RECOMPUTE_CSUM(flags))
+               skb_postpush_rcsum(skb, ptr, len);
+
        return 0;
 }
 
@@ -1525,8 +1526,12 @@ static u64 bpf_clone_redirect(u64 r1, u64 ifindex, u64 flags, u64 r4, u64 r5)
        if (unlikely(!skb2))
                return -ENOMEM;
 
-       if (BPF_IS_REDIRECT_INGRESS(flags))
+       if (BPF_IS_REDIRECT_INGRESS(flags)) {
+               if (skb_at_tc_ingress(skb2))
+                       skb_postpush_rcsum(skb2, skb_mac_header(skb2),
+                                          skb2->mac_len);
                return dev_forward_skb(dev, skb2);
+       }
 
        skb2->dev = dev;
        skb_sender_cpu_clear(skb2);
@@ -1569,8 +1574,12 @@ int skb_do_redirect(struct sk_buff *skb)
                return -EINVAL;
        }
 
-       if (BPF_IS_REDIRECT_INGRESS(ri->flags))
+       if (BPF_IS_REDIRECT_INGRESS(ri->flags)) {
+               if (skb_at_tc_ingress(skb))
+                       skb_postpush_rcsum(skb, skb_mac_header(skb),
+                                          skb->mac_len);
                return dev_forward_skb(dev, skb);
+       }
 
        skb->dev = dev;
        skb_sender_cpu_clear(skb);