On 07/12/12 08:20, Dave Taht wrote:
Also, some new unaligned exception traps have surfaced with ipv6. They
are not bad - only about 1200 on a 60 second rrul test with 24.5/5.5
ceroshaper, but enough to dramatically affect ipv6 latencies, and
affect fq_codel. (see attached)
I have managed to track down possible causes of these new traps now.
I've noticed that in the IPv6 stack in particular, there's a tendency to
access the flow label and traffic class fields by casting the header
structure to __be32*. Needless to say, this ends up producing a lot of
unaligned access traps.
There's a possible issue with setting the IPv6 ECN bits, but that should
also affect the Sugarland build too. There's also a few *(__be32*)
dereferences in net/ipv6/ip6_route.c and net/ipv6/ip6_tunnel.c that are
new to the 3.6 kernel. The attached patches should hopefully fix all of
the new traps.
--
Robert Bradley
diff --git a/include/net/inet_ecn.h b/include/net/inet_ecn.h
index 2fa1469..0b6a7fd 100644
--- a/include/net/inet_ecn.h
+++ b/include/net/inet_ecn.h
@@ -111,15 +111,21 @@ struct ipv6hdr;
static inline int IP6_ECN_set_ce(struct ipv6hdr *iph)
{
+ __be32 dsfield;
if (INET_ECN_is_not_ect(ipv6_get_dsfield(iph)))
return 0;
- *(__be32*)iph |= htonl(INET_ECN_CE << 20);
+ dsfield = __get_unaligned_cpu32((__be32*)iph);
+ dsfield |= htonl(INET_ECN_CE << 20);
+ __put_unaligned_cpu32(dsfield, (__be32*)iph);
return 1;
}
static inline void IP6_ECN_clear(struct ipv6hdr *iph)
{
- *(__be32*)iph &= ~htonl(INET_ECN_MASK << 20);
+ __be32 dsfield;
+ dsfield = __get_unaligned_cpu32((__be32*)iph);
+ dsfield &= ~htonl(INET_ECN_MASK << 20);
+ __put_unaligned_cpu32(dsfield, (__be32*)iph);
}
static inline void ipv6_copy_dscp(unsigned int dscp, struct ipv6hdr *inner)
diff --git a/net/ipv6/datagram.c b/net/ipv6/datagram.c
index 19e7aa0..6cbd2b7 100644
--- a/net/ipv6/datagram.c
+++ b/net/ipv6/datagram.c
@@ -492,7 +492,7 @@ int datagram_recv_ctl(struct sock *sk, struct msghdr *msg, struct sk_buff *skb)
put_cmsg(msg, SOL_IPV6, IPV6_TCLASS, sizeof(tclass), &tclass);
}
- if (np->rxopt.bits.rxflow && (*(__be32 *)nh & IPV6_FLOWINFO_MASK)) {
+ if (np->rxopt.bits.rxflow && (__get_unaligned_cpu32((__be32 *)nh) & IPV6_FLOWINFO_MASK)) {
__be32 flowinfo = __get_unaligned_cpu32((__be32 *)nh) & IPV6_FLOWINFO_MASK;
put_cmsg(msg, SOL_IPV6, IPV6_FLOWINFO, sizeof(flowinfo), &flowinfo);
}
diff --git a/net/ipv6/ip6_output.c b/net/ipv6/ip6_output.c
index 5b2d63e..b6689e3 100644
--- a/net/ipv6/ip6_output.c
+++ b/net/ipv6/ip6_output.c
@@ -221,7 +221,9 @@ int ip6_xmit(struct sock *sk, struct sk_buff *skb, struct flowi6 *fl6,
if (hlimit < 0)
hlimit = ip6_dst_hoplimit(dst);
- *(__be32 *)hdr = htonl(0x60000000 | (tclass << 20)) | fl6->flowlabel;
+ __put_unaligned_cpu32(htonl(0x60000000 | (tclass << 20))
+ | fl6->flowlabel,
+ (__be32*)hdr);
hdr->payload_len = htons(seg_len);
hdr->nexthdr = proto;
@@ -272,7 +274,7 @@ int ip6_nd_hdr(struct sock *sk, struct sk_buff *skb, struct net_device *dev,
skb_put(skb, sizeof(struct ipv6hdr));
hdr = ipv6_hdr(skb);
- *(__be32*)hdr = htonl(0x60000000);
+ __put_unaligned_cpu32(htonl(0x60000000), (__be32*)hdr);
hdr->payload_len = htons(len);
hdr->nexthdr = proto;
@@ -1635,8 +1637,9 @@ int ip6_push_pending_frames(struct sock *sk)
skb_reset_network_header(skb);
hdr = ipv6_hdr(skb);
- *(__be32*)hdr = fl6->flowlabel |
- htonl(0x60000000 | ((int)np->cork.tclass << 20));
+ __put_unaligned_cpu32(fl6->flowlabel |
+ htonl(0x60000000 | ((int)np->cork.tclass << 20)),
+ (__be32*)hdr);
hdr->hop_limit = np->cork.hop_limit;
hdr->nexthdr = proto;
diff --git a/net/ipv6/ip6_tunnel.c b/net/ipv6/ip6_tunnel.c
index 9a1d5fe..455ea7f 100644
--- a/net/ipv6/ip6_tunnel.c
+++ b/net/ipv6/ip6_tunnel.c
@@ -997,7 +997,7 @@ static int ip6_tnl_xmit2(struct sk_buff *skb,
skb_push(skb, sizeof(struct ipv6hdr));
skb_reset_network_header(skb);
ipv6h = ipv6_hdr(skb);
- *(__be32*)ipv6h = fl6->flowlabel | htonl(0x60000000);
+ __put_unaligned_cpu32(fl6->flowlabel | htonl(0x60000000), (__be32*)ipv6h);
dsfield = INET_ECN_encapsulate(0, dsfield);
ipv6_change_dsfield(ipv6h, ~INET_ECN_MASK, dsfield);
ipv6h->hop_limit = t->parms.hop_limit;
@@ -1103,9 +1103,9 @@ ip6ip6_tnl_xmit(struct sk_buff *skb, struct net_device *dev)
dsfield = ipv6_get_dsfield(ipv6h);
if (t->parms.flags & IP6_TNL_F_USE_ORIG_TCLASS)
- fl6.flowlabel |= (*(__be32 *) ipv6h & IPV6_TCLASS_MASK);
+ fl6.flowlabel |= (__get_unaligned_cpu32((__be32 *) ipv6h) & IPV6_TCLASS_MASK);
if (t->parms.flags & IP6_TNL_F_USE_ORIG_FLOWLABEL)
- fl6.flowlabel |= (*(__be32 *) ipv6h & IPV6_FLOWLABEL_MASK);
+ fl6.flowlabel |= (__get_unaligned_cpu32((__be32 *) ipv6h) & IPV6_FLOWLABEL_MASK);
if (t->parms.flags & IP6_TNL_F_USE_ORIG_FWMARK)
fl6.flowi6_mark = skb->mark;
diff --git a/net/ipv6/route.c b/net/ipv6/route.c
index 3e350eb..c2787a1 100644
--- a/net/ipv6/route.c
+++ b/net/ipv6/route.c
@@ -1117,7 +1117,7 @@ void ip6_update_pmtu(struct sk_buff *skb, struct net *net, __be32 mtu,
fl6.flowi6_flags = 0;
fl6.daddr = iph->daddr;
fl6.saddr = iph->saddr;
- fl6.flowlabel = (*(__be32 *) iph) & IPV6_FLOWINFO_MASK;
+ fl6.flowlabel = __get_unaligned_cpu32((__be32 *) iph) & IPV6_FLOWINFO_MASK;
dst = ip6_route_output(net, NULL, &fl6);
if (!dst->error)
@@ -1145,7 +1145,7 @@ void ip6_redirect(struct sk_buff *skb, struct net *net, int oif, u32 mark)
fl6.flowi6_flags = 0;
fl6.daddr = iph->daddr;
fl6.saddr = iph->saddr;
- fl6.flowlabel = (*(__be32 *) iph) & IPV6_FLOWINFO_MASK;
+ fl6.flowlabel = __get_unaligned_cpu32((__be32 *) iph) & IPV6_FLOWINFO_MASK;
dst = ip6_route_output(net, NULL, &fl6);
if (!dst->error)
_______________________________________________
Cerowrt-devel mailing list
[email protected]
https://lists.bufferbloat.net/listinfo/cerowrt-devel
