Skip to contentSkip to conversation

IPVS optimizations

PATCH In review Julian Anastasov 3 participants 20 messages Mar 21, 2013 – Mar 28, 2013

Subsystems (4)
lore
Share & tools
lore

Julian Anastasov

author maintainer
#0 lore raw

v3 cover IPVS optimizations

This is a first patchset for IPVS optimizations.
Another patchset will address the locking in schedulers and moving the global _bh disabling from LOCAL_OUT to all locks.

All patches are for net-next and Simon can
take them for ipvs-next.

The changes in this patchset eliminate locks
and dst refcnt operations from packet processing by using RCU. There are more details in the patches.

v3:
* in "ipvs: consolidate all dst checks on transmit in one place" preserve original skb dst even for local client, remove the rt_is_input_route and skb->dev check. Call update_pmtu only for local client by providing sk instead of skb.
* in "ipvs: optimize dst usage for real server" use rcu_dereference_protected for __ip_vs_dst_cache_reset instead of rcu_dereference_raw. Use the new skb_dst_set_noref_force func.
* in "ipvs: remove rs_lock by using RCU" prefer the port check in ip_vs_has_real_service
* "ipvs: convert locks used in persistence engines" needs only synchronize_rcu, not rcu_barrier, we do not use rcu callbacks

v2:
* use "net: add skb_dst_set_unref" instead of
"net: add dst_get_noref and refdst_ptr helpers"
* add "ipvs: no need to reroute anymore on DNAT over loopback"
* add "ipvs: do not use skb_share_check"
* add "ipvs: consolidate all dst checks on transmit in one place", so that we can avoid the refdst games in next patch
* after "ipvs: consolidate all dst checks on transmit in one place" "ipvs: optimize dst usage for real server" is simpler and uses the new skb_dst_set_unref function
* extend "ipvs: reorder keys in connection structure" with changes in ip_vs_ct_in_get
* fix "ipvs: avoid kmem_cache_zalloc in ip_vs_conn_new" to use new function ip_vs_addr_set, so that we reset all address fields that are used for hashing by hash_conntrack_raw

Julian Anastasov (15):
net: add skb_dst_set_noref_force
ipvs: avoid routing by TOS for real server
ipvs: prefer NETDEV_DOWN event to free cached dsts
ipvs: convert the IP_VS_XMIT macros to functions
ipvs: rename functions related to dst_cache reset
ipvs: no need to reroute anymore on DNAT over loopback
ipvs: do not use skb_share_check
ipvs: consolidate all dst checks on transmit in one place
ipvs: optimize dst usage for real server
ipvs: convert app locks
ipvs: remove rs_lock by using RCU
ipvs: convert locks used in persistence engines
ipvs: convert connection locking
ipvs: reorder keys in connection structure
ipvs: avoid kmem_cache_zalloc in ip_vs_conn_new

 include/linux/skbuff.h                |   35 ++-
 include/net/ip_vs.h                   |   71 ++-
 net/core/dst.c                        |    9 +-
 net/netfilter/ipvs/ip_vs_app.c        |   27 +-
 net/netfilter/ipvs/ip_vs_conn.c       |  271 +++++----
 net/netfilter/ipvs/ip_vs_core.c       |   16 +-
 net/netfilter/ipvs/ip_vs_ctl.c        |  143 +++---
 net/netfilter/ipvs/ip_vs_ftp.c        |    2 +
 net/netfilter/ipvs/ip_vs_pe.c         |   43 +-
 net/netfilter/ipvs/ip_vs_pe_sip.c     |    1 +
 net/netfilter/ipvs/ip_vs_proto_sctp.c |   18 +-
 net/netfilter/ipvs/ip_vs_proto_tcp.c  |   18 +-
 net/netfilter/ipvs/ip_vs_proto_udp.c  |   19 +-
 net/netfilter/ipvs/ip_vs_xmit.c       | 1046 ++++++++++++++-------------------
 14 files changed, 810 insertions(+), 909 deletions(-)
Signature · 1 line
1.7.3.4
author maintainer
#1 lore raw

PATCH v3 1/15 net: add skb_dst_set_noref_force

Rename skb_dst_set_noref to __skb_dst_set_noref and
add force flag as suggested by David Miller. The new wrapper skb_dst_set_noref_force will force dst entries that are not cached to be attached as skb dst without taking reference as long as provided dst is reclaimed after RCU grace period.

---
 include/linux/skbuff.h |   35 ++++++++++++++++++++++++++++++++++-
 net/core/dst.c         |    9 +++++----
 2 files changed, 39 insertions(+), 5 deletions(-)
include/linux/skbuff.h +34 -1
index 4974121..6cf3a46 100644
@@ -574,7 +574,40 @@ static inline void skb_dst_set(struct sk_buff *skb, struct dst_entry *dst)
 	skb->_skb_refdst = (unsigned long)dst;
 }
 
-extern void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst);
+extern void __skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst,
+				bool force);
+
+/**
+ * skb_dst_set_noref - sets skb dst, hopefully, without taking reference
+ * @skb: buffer
+ * @dst: dst entry
+ *
+ * Sets skb dst, assuming a reference was not taken on dst.
+ * If dst entry is cached, we do not take reference and dst_release
+ * will be avoided by refdst_drop. If dst entry is not cached, we take
+ * reference, so that last dst_release can destroy the dst immediately.
+ */
+static inline void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
+{
+	__skb_dst_set_noref(skb, dst, false);
+}
+
+/**
+ * skb_dst_set_noref_force - sets skb dst, without taking reference
+ * @skb: buffer
+ * @dst: dst entry
+ *
+ * Sets skb dst, assuming a reference was not taken on dst.
+ * No reference is taken and no dst_release will be called. While for
+ * cached dsts deferred reclaim is a basic feature, for entries that are
+ * not cached it is caller's job to guarantee that last dst_release for
+ * provided dst happens when nobody uses it, eg. after a RCU grace period.
+ */
+static inline void skb_dst_set_noref_force(struct sk_buff *skb,
+					   struct dst_entry *dst)
+{
+	__skb_dst_set_noref(skb, dst, true);
+}
 
 /**
  * skb_dst_is_noref - Test if skb dst isn't refcounted
net/core/dst.c +5 -4
index 35fd12f..df9cc81 100644
@@ -320,27 +320,28 @@ void __dst_destroy_metrics_generic(struct dst_entry *dst, unsigned long old)
 EXPORT_SYMBOL(__dst_destroy_metrics_generic);
 
 /**
- * skb_dst_set_noref - sets skb dst, without a reference
+ * __skb_dst_set_noref - sets skb dst, without a reference
  * @skb: buffer
  * @dst: dst entry
+ * @force: if force is set, use noref version even for DST_NOCACHE entries
  *
  * Sets skb dst, assuming a reference was not taken on dst
  * skb_dst_drop() should not dst_release() this dst
  */
-void skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst)
+void __skb_dst_set_noref(struct sk_buff *skb, struct dst_entry *dst, bool force)
 {
 	WARN_ON(!rcu_read_lock_held() && !rcu_read_lock_bh_held());
 	/* If dst not in cache, we must take a reference, because
 	 * dst_release() will destroy dst as soon as its refcount becomes zero
 	 */
-	if (unlikely(dst->flags & DST_NOCACHE)) {
+	if (unlikely((dst->flags & DST_NOCACHE) && !force)) {
 		dst_hold(dst);
 		skb_dst_set(skb, dst);
 	} else {
 		skb->_skb_refdst = (unsigned long)dst | SKB_DST_NOREF;
 	}
 }
-EXPORT_SYMBOL(skb_dst_set_noref);
+EXPORT_SYMBOL(__skb_dst_set_noref);
 
 /* Dirty hack. We did it in 2.2 (in __dst_free),
  * we have _very_ good reasons not to repeat
Signature · 1 line
1.7.3.4
author maintainer
#2 lore raw

PATCH v3 2/15 ipvs: avoid routing by TOS for real server

In reply to Julian Anastasov

Avoid replacing the cached route for real server
on every packet with different TOS. I doubt that routing by TOS for real server is used at all, so we should be better with such optimization.

---
 include/net/ip_vs.h             |    1 -
 net/netfilter/ipvs/ip_vs_xmit.c |   58 +++++++++++++++++----------------------
 2 files changed, 25 insertions(+), 34 deletions(-)
include/net/ip_vs.h +0 -1
index fce8e6b..2d14ba8 100644
@@ -753,7 +753,6 @@ struct ip_vs_dest {
 	/* for destination cache */
 	spinlock_t		dst_lock;	/* lock of dst_cache */
 	struct dst_entry	*dst_cache;	/* destination cache entry */
-	u32			dst_rtos;	/* RT_TOS(tos) for dst */
 	u32			dst_cookie;
 	union nf_inet_addr	dst_saddr;
 
net/netfilter/ipvs/ip_vs_xmit.c +25 -33

169 lines not on this page · raw message

Signature · 1 line
1.7.3.4
author maintainer
#3 lore raw

PATCH v3 3/15 ipvs: prefer NETDEV_DOWN event to free cached dsts

In reply to Julian Anastasov

The real server becomes unreachable on down event,
no need to wait device unregistration. Should help in releasing dsts early before dst->dev is replaced with lo.

Signed-off by: Hans Schillstrom <[email protected]>

---
 net/netfilter/ipvs/ip_vs_ctl.c |    8 +++-----
 1 files changed, 3 insertions(+), 5 deletions(-)
net/netfilter/ipvs/ip_vs_ctl.c +3 -5
index 9e2d1cc..1648454 100644
@@ -1512,10 +1512,8 @@ __ip_vs_dev_reset(struct ip_vs_dest *dest, struct net_device *dev)
 	spin_unlock_bh(&dest->dst_lock);
 
 }
-/*
- * Netdev event receiver
- * Currently only NETDEV_UNREGISTER is handled, i.e. if we hold a reference to
- * a device that is "unregister" it must be released.
+/* Netdev event receiver
+ * Currently only NETDEV_DOWN is handled to release refs to cached dsts
  */
 static int ip_vs_dst_event(struct notifier_block *this, unsigned long event,
 			    void *ptr)
@@ -1527,7 +1525,7 @@ static int ip_vs_dst_event(struct notifier_block *this, unsigned long event,
 	struct ip_vs_dest *dest;
 	unsigned int idx;
 
-	if (event != NETDEV_UNREGISTER || !ipvs)
+	if (event != NETDEV_DOWN || !ipvs)
 		return NOTIFY_DONE;
 	IP_VS_DBG(3, "%s() dev=%s\n", __func__, dev->name);
 	EnterFunction(2);
Signature · 1 line
1.7.3.4
author maintainer
#4 lore raw

PATCH v3 4/15 ipvs: convert the IP_VS_XMIT macros to functions

In reply to Julian Anastasov

It was a bad idea to hide return statements in macros.

---
 net/netfilter/ipvs/ip_vs_xmit.c |  134 +++++++++++++++++++++------------------
 1 files changed, 72 insertions(+), 62 deletions(-)
net/netfilter/ipvs/ip_vs_xmit.c +71 -60
index 4b0bd15..7cd7c61 100644
@@ -376,45 +376,59 @@ ip_vs_dst_reset(struct ip_vs_dest *dest)
 	dest->dst_saddr.ip = 0;
 }
 
-#define IP_VS_XMIT_TUNNEL(skb, cp)				\
-({								\
-	int __ret = NF_ACCEPT;					\
-								\
-	(skb)->ipvs_property = 1;				\
-	if (unlikely((cp)->flags & IP_VS_CONN_F_NFCT))		\
-		__ret = ip_vs_confirm_conntrack(skb);		\
-	if (__ret == NF_ACCEPT) {				\
-		nf_reset(skb);					\
-		skb_forward_csum(skb);				\
-	}							\
-	__ret;							\
-})
-
-#define IP_VS_XMIT_NAT(pf, skb, cp, local)		\
-do {							\
-	(skb)->ipvs_property = 1;			\
-	if (likely(!((cp)->flags & IP_VS_CONN_F_NFCT)))	\
-		ip_vs_notrack(skb);			\
-	else						\
-		ip_vs_update_conntrack(skb, cp, 1);	\
-	if (local)					\
-		return NF_ACCEPT;			\
-	skb_forward_csum(skb);				\
-	NF_HOOK(pf, NF_INET_LOCAL_OUT, (skb), NULL,	\
-		skb_dst(skb)->dev, dst_output);		\
-} while (0)
-
-#define IP_VS_XMIT(pf, skb, cp, local)			\
-do {							\
-	(skb)->ipvs_property = 1;			\
-	if (likely(!((cp)->flags & IP_VS_CONN_F_NFCT)))	\
-		ip_vs_notrack(skb);			\
-	if (local)					\
-		return NF_ACCEPT;			\
-	skb_forward_csum(skb);				\
-	NF_HOOK(pf, NF_INET_LOCAL_OUT, (skb), NULL,	\
-		skb_dst(skb)->dev, dst_output);		\
-} while (0)
+/* return NF_ACCEPT to allow forwarding or other NF_xxx on error */
+static inline int ip_vs_tunnel_xmit_prepare(struct sk_buff *skb,
+					    struct ip_vs_conn *cp)
+{
+	int ret = NF_ACCEPT;
+
+	skb->ipvs_property = 1;
+	if (unlikely(cp->flags & IP_VS_CONN_F_NFCT))
+		ret = ip_vs_confirm_conntrack(skb);
+	if (ret == NF_ACCEPT) {
+		nf_reset(skb);
+		skb_forward_csum(skb);
+	}
+	return ret;
+}
+
+/* return NF_STOLEN (sent) or NF_ACCEPT if local=1 (not sent) */
+static inline int ip_vs_nat_send_or_cont(int pf, struct sk_buff *skb,
+					 struct ip_vs_conn *cp, int local)
+{
+	int ret = NF_STOLEN;
+
+	skb->ipvs_property = 1;
+	if (likely(!(cp->flags & IP_VS_CONN_F_NFCT)))
+		ip_vs_notrack(skb);
+	else
+		ip_vs_update_conntrack(skb, cp, 1);
+	if (!local) {
+		skb_forward_csum(skb);
+		NF_HOOK(pf, NF_INET_LOCAL_OUT, skb, NULL, skb_dst(skb)->dev,
+			dst_output);
+	} else
+		ret = NF_ACCEPT;
+	return ret;
+}
+
+/* return NF_STOLEN (sent) or NF_ACCEPT if local=1 (not sent) */
+static inline int ip_vs_send_or_cont(int pf, struct sk_buff *skb,
+				     struct ip_vs_conn *cp, int local)
+{
+	int ret = NF_STOLEN;
+
+	skb->ipvs_property = 1;
+	if (likely(!(cp->flags & IP_VS_CONN_F_NFCT)))
+		ip_vs_notrack(skb);
+	if (!local) {
+		skb_forward_csum(skb);
+		NF_HOOK(pf, NF_INET_LOCAL_OUT, skb, NULL, skb_dst(skb)->dev,
+			dst_output);
+	} else
+		ret = NF_ACCEPT;
+	return ret;
+}
 
 
 /*
@@ -425,7 +439,7 @@ ip_vs_null_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		struct ip_vs_protocol *pp, struct ip_vs_iphdr *ipvsh)
 {
 	/* we do not touch skb and do not need pskb ptr */
-	IP_VS_XMIT(NFPROTO_IPV4, skb, cp, 1);
+	return ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 1);
 }
 
 
@@ -476,7 +490,7 @@ ip_vs_bypass_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT(NFPROTO_IPV4, skb, cp, 0);
+	ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 0);
 
 	LeaveFunction(10);
 	return NF_STOLEN;
@@ -537,7 +551,7 @@ ip_vs_bypass_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT(NFPROTO_IPV6, skb, cp, 0);
+	ip_vs_send_or_cont(NFPROTO_IPV6, skb, cp, 0);
 
 	LeaveFunction(10);
 	return NF_STOLEN;
@@ -562,7 +576,7 @@ ip_vs_nat_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	struct rtable *rt;		/* Route to the other host */
 	int mtu;
 	struct iphdr *iph = ip_hdr(skb);
-	int local;
+	int local, rc;
 
 	EnterFunction(10);
 
@@ -655,10 +669,10 @@ ip_vs_nat_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT_NAT(NFPROTO_IPV4, skb, cp, local);
+	rc = ip_vs_nat_send_or_cont(NFPROTO_IPV4, skb, cp, local);
 
 	LeaveFunction(10);
-	return NF_STOLEN;
+	return rc;
 
   tx_error_icmp:
 	dst_link_failure(skb);
@@ -678,7 +692,7 @@ ip_vs_nat_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 {
 	struct rt6_info *rt;		/* Route to the other host */
 	int mtu;
-	int local;
+	int local, rc;
 
 	EnterFunction(10);
 
@@ -771,10 +785,10 @@ ip_vs_nat_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT_NAT(NFPROTO_IPV6, skb, cp, local);
+	rc = ip_vs_nat_send_or_cont(NFPROTO_IPV6, skb, cp, local);
 
 	LeaveFunction(10);
-	return NF_STOLEN;
+	return rc;
 
 tx_error_icmp:
 	dst_link_failure(skb);
@@ -833,7 +847,7 @@ ip_vs_tunnel_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error_icmp;
 	if (rt->rt_flags & RTCF_LOCAL) {
 		ip_rt_put(rt);
-		IP_VS_XMIT(NFPROTO_IPV4, skb, cp, 1);
+		return ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 1);
 	}
 
 	tdev = rt->dst.dev;
@@ -905,7 +919,7 @@ ip_vs_tunnel_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	ret = IP_VS_XMIT_TUNNEL(skb, cp);
+	ret = ip_vs_tunnel_xmit_prepare(skb, cp);
 	if (ret == NF_ACCEPT)
 		ip_local_out(skb);
 	else if (ret == NF_DROP)
@@ -948,7 +962,7 @@ ip_vs_tunnel_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error_icmp;
 	if (__ip_vs_is_local_route6(rt)) {
 		dst_release(&rt->dst);
-		IP_VS_XMIT(NFPROTO_IPV6, skb, cp, 1);
+		return ip_vs_send_or_cont(NFPROTO_IPV6, skb, cp, 1);
 	}
 
 	tdev = rt->dst.dev;
@@ -1023,7 +1037,7 @@ ip_vs_tunnel_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	ret = IP_VS_XMIT_TUNNEL(skb, cp);
+	ret = ip_vs_tunnel_xmit_prepare(skb, cp);
 	if (ret == NF_ACCEPT)
 		ip6_local_out(skb);
 	else if (ret == NF_DROP)
@@ -1067,7 +1081,7 @@ ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error_icmp;
 	if (rt->rt_flags & RTCF_LOCAL) {
 		ip_rt_put(rt);
-		IP_VS_XMIT(NFPROTO_IPV4, skb, cp, 1);
+		return ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 1);
 	}
 
 	/* MTU checking */
@@ -1097,7 +1111,7 @@ ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT(NFPROTO_IPV4, skb, cp, 0);
+	ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 0);
 
 	LeaveFunction(10);
 	return NF_STOLEN;
@@ -1126,7 +1140,7 @@ ip_vs_dr_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error_icmp;
 	if (__ip_vs_is_local_route6(rt)) {
 		dst_release(&rt->dst);
-		IP_VS_XMIT(NFPROTO_IPV6, skb, cp, 1);
+		return ip_vs_send_or_cont(NFPROTO_IPV6, skb, cp, 1);
 	}
 
 	/* MTU checking */
@@ -1162,7 +1176,7 @@ ip_vs_dr_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT(NFPROTO_IPV6, skb, cp, 0);
+	ip_vs_send_or_cont(NFPROTO_IPV6, skb, cp, 0);
 
 	LeaveFunction(10);
 	return NF_STOLEN;
@@ -1283,9 +1297,7 @@ ip_vs_icmp_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT_NAT(NFPROTO_IPV4, skb, cp, local);
-
-	rc = NF_STOLEN;
+	rc = ip_vs_nat_send_or_cont(NFPROTO_IPV4, skb, cp, local);
 	goto out;
 
   tx_error_icmp:
@@ -1404,9 +1416,7 @@ ip_vs_icmp_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
-	IP_VS_XMIT_NAT(NFPROTO_IPV6, skb, cp, local);
author maintainer
#5 lore raw

PATCH v3 5/15 ipvs: rename functions related to dst_cache reset

In reply to Julian Anastasov

Move and give better names to two functions:

- ip_vs_dst_reset to __ip_vs_dst_cache_reset
- __ip_vs_dev_reset to ip_vs_forget_dev

---
 include/net/ip_vs.h             |    1 -
 net/netfilter/ipvs/ip_vs_ctl.c  |   34 ++++++++++++++++++++++------------
 net/netfilter/ipvs/ip_vs_xmit.c |   14 --------------
 3 files changed, 22 insertions(+), 27 deletions(-)
include/net/ip_vs.h +0 -1
index 2d14ba8..c448d3c 100644
@@ -1427,7 +1427,6 @@ extern int ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 extern int ip_vs_icmp_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 			   struct ip_vs_protocol *pp, int offset,
 			   unsigned int hooknum, struct ip_vs_iphdr *iph);
-extern void ip_vs_dst_reset(struct ip_vs_dest *dest);
 
 #ifdef CONFIG_IP_VS_IPV6
 extern int ip_vs_bypass_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
net/netfilter/ipvs/ip_vs_ctl.c +22 -12
index 1648454..b545549 100644
@@ -639,6 +639,17 @@ struct ip_vs_dest *ip_vs_find_dest(struct net  *net, int af,
 	return dest;
 }
 
+/* Release dst_cache for dest in user context */
+static void __ip_vs_dst_cache_reset(struct ip_vs_dest *dest)
+{
+	struct dst_entry *old_dst;
+
+	old_dst = dest->dst_cache;
+	dest->dst_cache = NULL;
+	dst_release(old_dst);
+	dest->dst_saddr.ip = 0;
+}
+
 /*
  *  Lookup dest by {svc,addr,port} in the destination trash.
  *  The destination trash is used to hold the destinations that are removed
@@ -688,7 +699,7 @@ ip_vs_trash_get_dest(struct ip_vs_service *svc, const union nf_inet_addr *daddr,
 				      IP_VS_DBG_ADDR(svc->af, &dest->addr),
 				      ntohs(dest->port));
 			list_del(&dest->n_list);
-			ip_vs_dst_reset(dest);
+			__ip_vs_dst_cache_reset(dest);
 			__ip_vs_unbind_svc(dest);
 			free_percpu(dest->stats.cpustats);
 			kfree(dest);
@@ -715,7 +726,7 @@ static void ip_vs_trash_cleanup(struct net *net)
 
 	list_for_each_entry_safe(dest, nxt, &ipvs->dest_trash, n_list) {
 		list_del(&dest->n_list);
-		ip_vs_dst_reset(dest);
+		__ip_vs_dst_cache_reset(dest);
 		__ip_vs_unbind_svc(dest);
 		free_percpu(dest->stats.cpustats);
 		kfree(dest);
@@ -809,7 +820,7 @@ __ip_vs_update_dest(struct ip_vs_service *svc, struct ip_vs_dest *dest,
 	dest->l_threshold = udest->l_threshold;
 
 	spin_lock_bh(&dest->dst_lock);
-	ip_vs_dst_reset(dest);
+	__ip_vs_dst_cache_reset(dest);
 	spin_unlock_bh(&dest->dst_lock);
 
 	if (add)
@@ -1035,7 +1046,7 @@ static void __ip_vs_del_dest(struct net *net, struct ip_vs_dest *dest)
 			      dest->vfwmark,
 			      IP_VS_DBG_ADDR(dest->af, &dest->addr),
 			      ntohs(dest->port));
-		ip_vs_dst_reset(dest);
+		__ip_vs_dst_cache_reset(dest);
 		/* simply decrease svc->refcnt here, let the caller check
 		   and release the service if nobody refers to it.
 		   Only user context can release destination and service,
@@ -1494,11 +1505,10 @@ void ip_vs_service_net_cleanup(struct net *net)
 	mutex_unlock(&__ip_vs_mutex);
 	LeaveFunction(2);
 }
-/*
- * Release dst hold by dst_cache
- */
+
+/* Put all references for device (dst_cache) */
 static inline void
-__ip_vs_dev_reset(struct ip_vs_dest *dest, struct net_device *dev)
+ip_vs_forget_dev(struct ip_vs_dest *dest, struct net_device *dev)
 {
 	spin_lock_bh(&dest->dst_lock);
 	if (dest->dst_cache && dest->dst_cache->dev == dev) {
@@ -1507,7 +1517,7 @@ __ip_vs_dev_reset(struct ip_vs_dest *dest, struct net_device *dev)
 			      IP_VS_DBG_ADDR(dest->af, &dest->addr),
 			      ntohs(dest->port),
 			      atomic_read(&dest->refcnt));
-		ip_vs_dst_reset(dest);
+		__ip_vs_dst_cache_reset(dest);
 	}
 	spin_unlock_bh(&dest->dst_lock);
 
@@ -1535,7 +1545,7 @@ static int ip_vs_dst_event(struct notifier_block *this, unsigned long event,
 			if (net_eq(svc->net, net)) {
 				list_for_each_entry(dest, &svc->destinations,
 						    n_list) {
-					__ip_vs_dev_reset(dest, dev);
+					ip_vs_forget_dev(dest, dev);
 				}
 			}
 		}
@@ -1544,7 +1554,7 @@ static int ip_vs_dst_event(struct notifier_block *this, unsigned long event,
 			if (net_eq(svc->net, net)) {
 				list_for_each_entry(dest, &svc->destinations,
 						    n_list) {
-					__ip_vs_dev_reset(dest, dev);
+					ip_vs_forget_dev(dest, dev);
 				}
 			}
 
@@ -1552,7 +1562,7 @@ static int ip_vs_dst_event(struct notifier_block *this, unsigned long event,
 	}
 
 	list_for_each_entry(dest, &ipvs->dest_trash, n_list) {
-		__ip_vs_dev_reset(dest, dev);
+		ip_vs_forget_dev(dest, dev);
 	}
 	mutex_unlock(&__ip_vs_mutex);
 	LeaveFunction(2);
net/netfilter/ipvs/ip_vs_xmit.c +0 -13
index 7cd7c61..6448a2e 100644
@@ -362,20 +362,6 @@ __ip_vs_get_out_rt_v6(struct sk_buff *skb, struct ip_vs_dest *dest,
 #endif
 
 
-/*
- *	Release dest->dst_cache before a dest is removed
- */
-void
-ip_vs_dst_reset(struct ip_vs_dest *dest)
-{
-	struct dst_entry *old_dst;
-
-	old_dst = dest->dst_cache;
-	dest->dst_cache = NULL;
-	dst_release(old_dst);
-	dest->dst_saddr.ip = 0;
-}
author maintainer
#6 lore raw

PATCH v3 6/15 ipvs: no need to reroute anymore on DNAT over loopback

In reply to Julian Anastasov
	After commit 70e7341673 (ipv4: Show that ip_send_reply()

is purely unicast routine.) we do not need to reroute DNAT-ed traffic over loopback because reply uses iph daddr and not rt_spec_dst.

---
 net/netfilter/ipvs/ip_vs_xmit.c |   58 +-------------------------------------
 1 files changed, 2 insertions(+), 56 deletions(-)
net/netfilter/ipvs/ip_vs_xmit.c +2 -56
index 6448a2e..c942d36 100644
@@ -207,44 +207,6 @@ __ip_vs_get_out_rt(struct sk_buff *skb, struct ip_vs_dest *dest,
 	return rt;
 }
 
-/* Reroute packet to local IPv4 stack after DNAT */
-static int
-__ip_vs_reroute_locally(struct sk_buff *skb)
-{
-	struct rtable *rt = skb_rtable(skb);
-	struct net_device *dev = rt->dst.dev;
-	struct net *net = dev_net(dev);
-	struct iphdr *iph = ip_hdr(skb);
-
-	if (rt_is_input_route(rt)) {
-		unsigned long orefdst = skb->_skb_refdst;
-
-		if (ip_route_input(skb, iph->daddr, iph->saddr,
-				   iph->tos, skb->dev))
-			return 0;
-		refdst_drop(orefdst);
-	} else {
-		struct flowi4 fl4 = {
-			.daddr = iph->daddr,
-			.saddr = iph->saddr,
-			.flowi4_tos = RT_TOS(iph->tos),
-			.flowi4_mark = skb->mark,
-		};
-
-		rt = ip_route_output_key(net, &fl4);
-		if (IS_ERR(rt))
-			return 0;
-		if (!(rt->rt_flags & RTCF_LOCAL)) {
-			ip_rt_put(rt);
-			return 0;
-		}
-		/* Drop old route. */
-		skb_dst_drop(skb);
-		skb_dst_set(skb, &rt->dst);
-	}
-	return 1;
-}
-
 #ifdef CONFIG_IP_VS_IPV6
 
 static inline int __ip_vs_is_local_route6(struct rt6_info *rt)
@@ -635,16 +597,8 @@ ip_vs_nat_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		/* drop old route */
 		skb_dst_drop(skb);
 		skb_dst_set(skb, &rt->dst);
-	} else {
+	} else
 		ip_rt_put(rt);
-		/*
-		 * Some IPv4 replies get local address from routes,
-		 * not from iph, so while we DNAT after routing
-		 * we need this second input/output route.
-		 */
-		if (!__ip_vs_reroute_locally(skb))
-			goto tx_error;
-	}
 
 	IP_VS_DBG_PKT(10, AF_INET, pp, skb, 0, "After DNAT");
 
@@ -1269,16 +1223,8 @@ ip_vs_icmp_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		/* drop the old route when skb is not shared */
 		skb_dst_drop(skb);
 		skb_dst_set(skb, &rt->dst);
-	} else {
+	} else
 		ip_rt_put(rt);
-		/*
-		 * Some IPv4 replies get local address from routes,
-		 * not from iph, so while we DNAT after routing
-		 * we need this second input/output route.
-		 */
-		if (!__ip_vs_reroute_locally(skb))
-			goto tx_error;
-	}
 
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
Signature · 1 line
1.7.3.4
author maintainer
#7 lore raw

PATCH v3 7/15 ipvs: do not use skb_share_check

In reply to Julian Anastasov

We run in contexts like ip_rcv, ipv6_rcv, br_handle_frame,
do not expect shared skbs.

---
 net/netfilter/ipvs/ip_vs_xmit.c |   45 +-------------------------------------
 1 files changed, 2 insertions(+), 43 deletions(-)
net/netfilter/ipvs/ip_vs_xmit.c +2 -40
index c942d36..e6e2d3f 100644
@@ -421,14 +421,6 @@ ip_vs_bypass_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error;
 	}
 
-	/*
-	 * Call ip_send_check because we are not sure it is called
-	 * after ip_defrag. Is copy-on-write needed?
-	 */
-	if (unlikely((skb = skb_share_check(skb, GFP_ATOMIC)) == NULL)) {
-		ip_rt_put(rt);
-		return NF_STOLEN;
-	}
 	ip_send_check(ip_hdr(skb));
 
 	/* drop old route */
@@ -482,16 +474,6 @@ ip_vs_bypass_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error;
 	}
 
-	/*
-	 * Call ip_send_check because we are not sure it is called
-	 * after ip_defrag. Is copy-on-write needed?
-	 */
-	skb = skb_share_check(skb, GFP_ATOMIC);
-	if (unlikely(skb == NULL)) {
-		dst_release(&rt->dst);
-		return NF_STOLEN;
-	}
-
 	/* drop old route */
 	skb_dst_drop(skb);
 	skb_dst_set(skb, &rt->dst);
@@ -708,7 +690,6 @@ ip_vs_nat_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	ipv6_hdr(skb)->daddr = cp->daddr.in6;
 
 	if (!local || !skb->dev) {
-		/* drop the old route when skb is not shared */
 		skb_dst_drop(skb);
 		skb_dst_set(skb, &rt->dst);
 	} else {
@@ -814,8 +795,7 @@ ip_vs_tunnel_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	 */
 	max_headroom = LL_RESERVED_SPACE(tdev) + sizeof(struct iphdr);
 
-	if (skb_headroom(skb) < max_headroom
-	    || skb_cloned(skb) || skb_shared(skb)) {
+	if (skb_headroom(skb) < max_headroom || skb_cloned(skb)) {
 		struct sk_buff *new_skb =
 			skb_realloc_headroom(skb, max_headroom);
 		if (!new_skb) {
@@ -935,8 +915,7 @@ ip_vs_tunnel_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	 */
 	max_headroom = LL_RESERVED_SPACE(tdev) + sizeof(struct ipv6hdr);
 
-	if (skb_headroom(skb) < max_headroom
-	    || skb_cloned(skb) || skb_shared(skb)) {
+	if (skb_headroom(skb) < max_headroom || skb_cloned(skb)) {
 		struct sk_buff *new_skb =
 			skb_realloc_headroom(skb, max_headroom);
 		if (!new_skb) {
@@ -1034,14 +1013,6 @@ ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error;
 	}
 
-	/*
-	 * Call ip_send_check because we are not sure it is called
-	 * after ip_defrag. Is copy-on-write needed?
-	 */
-	if (unlikely((skb = skb_share_check(skb, GFP_ATOMIC)) == NULL)) {
-		ip_rt_put(rt);
-		return NF_STOLEN;
-	}
 	ip_send_check(ip_hdr(skb));
 
 	/* drop old route */
@@ -1099,16 +1070,6 @@ ip_vs_dr_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 		goto tx_error;
 	}
 
-	/*
-	 * Call ip_send_check because we are not sure it is called
-	 * after ip_defrag. Is copy-on-write needed?
-	 */
-	skb = skb_share_check(skb, GFP_ATOMIC);
-	if (unlikely(skb == NULL)) {
-		dst_release(&rt->dst);
-		return NF_STOLEN;
-	}
author maintainer
#8 lore raw

PATCH v3 8/15 ipvs: consolidate all dst checks on transmit in one place

In reply to Julian Anastasov

Consolidate the PMTU checks, ICMP sending and
skb_dst modification in __ip_vs_get_out_rt and __ip_vs_get_out_rt_v6. Now skb_dst is changed early to simplify the transmitters.

Make sure update_pmtu is called only for
local clients.

---
 net/netfilter/ipvs/ip_vs_xmit.c |  627 +++++++++++++++------------------------
 1 files changed, 235 insertions(+), 392 deletions(-)
net/netfilter/ipvs/ip_vs_xmit.c +235 -386
index e6e2d3f..603eb8a 100644
@@ -51,6 +51,7 @@ enum {
 				      */
 	IP_VS_RT_MODE_CONNECT	= 8, /* Always bind route to saddr */
 	IP_VS_RT_MODE_KNOWN_NH	= 16,/* Route via remote addr */
+	IP_VS_RT_MODE_TUNNEL	= 32,/* Tunnel mode */
 };
 
 /*
@@ -137,13 +138,17 @@ retry:
 }
 
 /* Get route to destination or remote server */
-static struct rtable *
+static int
 __ip_vs_get_out_rt(struct sk_buff *skb, struct ip_vs_dest *dest,
 		   __be32 daddr, int rt_mode, __be32 *ret_saddr)
 {
 	struct net *net = dev_net(skb_dst(skb)->dev);
+	struct netns_ipvs *ipvs = net_ipvs(net);
 	struct rtable *rt;			/* Route to the other host */
 	struct rtable *ort;			/* Original route */
+	struct iphdr *iph;
+	__be16 df;
+	int mtu;
 	int local;
 
 	if (dest) {
@@ -154,7 +159,7 @@ __ip_vs_get_out_rt(struct sk_buff *skb, struct ip_vs_dest *dest,
 					      &dest->dst_saddr.ip);
 			if (!rt) {
 				spin_unlock(&dest->dst_lock);
-				return NULL;
+				goto err_unreach;
 			}
 			__ip_vs_dst_set(dest, dst_clone(&rt->dst), 0);
 			IP_VS_DBG(10, "new dst %pI4, src %pI4, refcnt=%d\n",
@@ -174,37 +179,78 @@ __ip_vs_get_out_rt(struct sk_buff *skb, struct ip_vs_dest *dest,
 		rt_mode &= ~IP_VS_RT_MODE_CONNECT;
 		rt = do_output_route4(net, daddr, rt_mode, &saddr);
 		if (!rt)
-			return NULL;
+			goto err_unreach;
 		if (ret_saddr)
 			*ret_saddr = saddr;
 	}
 
-	local = rt->rt_flags & RTCF_LOCAL;
+	local = (rt->rt_flags & RTCF_LOCAL) ? 1 : 0;
 	if (!((local ? IP_VS_RT_MODE_LOCAL : IP_VS_RT_MODE_NON_LOCAL) &
 	      rt_mode)) {
 		IP_VS_DBG_RL("Stopping traffic to %s address, dest: %pI4\n",
 			     (rt->rt_flags & RTCF_LOCAL) ?
 			     "local":"non-local", &daddr);
-		ip_rt_put(rt);
-		return NULL;
+		goto err_put;
 	}
-	if (local && !(rt_mode & IP_VS_RT_MODE_RDR) &&
-	    !((ort = skb_rtable(skb)) && ort->rt_flags & RTCF_LOCAL)) {
-		IP_VS_DBG_RL("Redirect from non-local address %pI4 to local "
-			     "requires NAT method, dest: %pI4\n",
-			     &ip_hdr(skb)->daddr, &daddr);
+	iph = ip_hdr(skb);
+	if (likely(!local)) {
+		if (unlikely(ipv4_is_loopback(iph->saddr))) {
+			IP_VS_DBG_RL("Stopping traffic from loopback address "
+				     "%pI4 to non-local address, dest: %pI4\n",
+				     &iph->saddr, &daddr);
+			goto err_put;
+		}
+	} else {
+		ort = skb_rtable(skb);
+		if (!(rt_mode & IP_VS_RT_MODE_RDR) &&
+		    !(ort->rt_flags & RTCF_LOCAL)) {
+			IP_VS_DBG_RL("Redirect from non-local address %pI4 to "
+				     "local requires NAT method, dest: %pI4\n",
+				     &iph->daddr, &daddr);
+			goto err_put;
+		}
+		/* skb to local stack, preserve old route */
 		ip_rt_put(rt);
-		return NULL;
+		return local;
 	}
-	if (unlikely(!local && ipv4_is_loopback(ip_hdr(skb)->saddr))) {
-		IP_VS_DBG_RL("Stopping traffic from loopback address %pI4 "
-			     "to non-local address, dest: %pI4\n",
-			     &ip_hdr(skb)->saddr, &daddr);
-		ip_rt_put(rt);
-		return NULL;
+
+	if (likely(!(rt_mode & IP_VS_RT_MODE_TUNNEL))) {
+		mtu = dst_mtu(&rt->dst);
+		df = iph->frag_off & htons(IP_DF);
+	} else {
+		struct sock *sk = skb->sk;
+
+		mtu = dst_mtu(&rt->dst) - sizeof(struct iphdr);
+		if (mtu < 68) {
+			IP_VS_DBG_RL("%s(): mtu less than 68\n", __func__);
+			goto err_put;
+		}
+		ort = skb_rtable(skb);
+		if (!skb->dev && sk && sk->sk_state != TCP_TIME_WAIT)
+			ort->dst.ops->update_pmtu(&ort->dst, sk, NULL, mtu);
+		/* MTU check allowed? */
+		df = sysctl_pmtu_disc(ipvs) ? iph->frag_off & htons(IP_DF) : 0;
 	}
 
-	return rt;
+	/* MTU checking */
+	if (unlikely(df && skb->len > mtu && !skb_is_gso(skb))) {
+		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
+		IP_VS_DBG(1, "frag needed for %pI4\n", &iph->saddr);
+		goto err_put;
+	}
+
+	skb_dst_drop(skb);
+	skb_dst_set(skb, &rt->dst);
+
+	return local;
+
+err_put:
+	ip_rt_put(rt);
+	return -1;
+
+err_unreach:
+	dst_link_failure(skb);
+	return -1;
 }
 
 #ifdef CONFIG_IP_VS_IPV6
@@ -251,15 +297,16 @@ out_err:
 /*
  * Get route to destination or remote server
  */
-static struct rt6_info *
+static int
 __ip_vs_get_out_rt_v6(struct sk_buff *skb, struct ip_vs_dest *dest,
 		      struct in6_addr *daddr, struct in6_addr *ret_saddr,
-		      int do_xfrm, int rt_mode)
+		      struct ip_vs_iphdr *ipvsh, int do_xfrm, int rt_mode)
 {
 	struct net *net = dev_net(skb_dst(skb)->dev);
 	struct rt6_info *rt;			/* Route to the other host */
 	struct rt6_info *ort;			/* Original route */
 	struct dst_entry *dst;
+	int mtu;
 	int local;
 
 	if (dest) {
@@ -273,7 +320,7 @@ __ip_vs_get_out_rt_v6(struct sk_buff *skb, struct ip_vs_dest *dest,
 						      do_xfrm);
 			if (!dst) {
 				spin_unlock(&dest->dst_lock);
-				return NULL;
+				goto err_unreach;
 			}
 			rt = (struct rt6_info *) dst;
 			cookie = rt->rt6i_node ? rt->rt6i_node->fn_sernum : 0;
@@ -288,7 +335,7 @@ __ip_vs_get_out_rt_v6(struct sk_buff *skb, struct ip_vs_dest *dest,
 	} else {
 		dst = __ip_vs_route_output_v6(net, daddr, ret_saddr, do_xfrm);
 		if (!dst)
-			return NULL;
+			goto err_unreach;
 		rt = (struct rt6_info *) dst;
 	}
 
@@ -297,29 +344,72 @@ __ip_vs_get_out_rt_v6(struct sk_buff *skb, struct ip_vs_dest *dest,
 	      rt_mode)) {
 		IP_VS_DBG_RL("Stopping traffic to %s address, dest: %pI6c\n",
 			     local ? "local":"non-local", daddr);
-		dst_release(&rt->dst);
-		return NULL;
+		goto err_put;
 	}
-	if (local && !(rt_mode & IP_VS_RT_MODE_RDR) &&
-	    !((ort = (struct rt6_info *) skb_dst(skb)) &&
-	      __ip_vs_is_local_route6(ort))) {
-		IP_VS_DBG_RL("Redirect from non-local address %pI6c to local "
-			     "requires NAT method, dest: %pI6c\n",
-			     &ipv6_hdr(skb)->daddr, daddr);
+	if (likely(!local)) {
+		if (unlikely((!skb->dev || skb->dev->flags & IFF_LOOPBACK) &&
+			     ipv6_addr_type(&ipv6_hdr(skb)->saddr) &
+					    IPV6_ADDR_LOOPBACK)) {
+			IP_VS_DBG_RL("Stopping traffic from loopback address "
+				     "%pI6c to non-local address, "
+				     "dest: %pI6c\n",
+				     &ipv6_hdr(skb)->saddr, daddr);
+			goto err_put;
+		}
+	} else {
+		ort = (struct rt6_info *) skb_dst(skb);
+		if (!(rt_mode & IP_VS_RT_MODE_RDR) &&
+		    !__ip_vs_is_local_route6(ort)) {
+			IP_VS_DBG_RL("Redirect from non-local address %pI6c "
+				     "to local requires NAT method, "
+				     "dest: %pI6c\n",
+				     &ipv6_hdr(skb)->daddr, daddr);
+			goto err_put;
+		}
+		/* skb to local stack, preserve old route */
 		dst_release(&rt->dst);
-		return NULL;
+		return local;
 	}
-	if (unlikely(!local && (!skb->dev || skb->dev->flags & IFF_LOOPBACK) &&
-		     ipv6_addr_type(&ipv6_hdr(skb)->saddr) &
-				    IPV6_ADDR_LOOPBACK)) {
-		IP_VS_DBG_RL("Stopping traffic from loopback address %pI6c "
-			     "to non-local address, dest: %pI6c\n",
-			     &ipv6_hdr(skb)->saddr, daddr);
-		dst_release(&rt->dst);
-		return NULL;
+
+	/* MTU checking */
+	if (likely(!(rt_mode & IP_VS_RT_MODE_TUNNEL)))
+		mtu = dst_mtu(&rt->dst);
+	else {
+		struct sock *sk = skb->sk;
+
+		mtu = dst_mtu(&rt->dst) - sizeof(struct ipv6hdr);
+		if (mtu < IPV6_MIN_MTU) {
+			IP_VS_DBG_RL("%s(): mtu less than %d\n", __func__,
+				     IPV6_MIN_MTU);
+			goto err_put;
+		}
+		ort = (struct rt6_info *) skb_dst(skb);
+		if (!skb->dev && sk && sk->sk_state != TCP_TIME_WAIT)
+			ort->dst.ops->update_pmtu(&ort->dst, sk, NULL, mtu);
 	}
 
-	return rt;
+	if (unlikely(__mtu_check_toobig_v6(skb, mtu))) {
+		if (!skb->dev)
+			skb->dev = net->loopback_dev;
+		/* only send ICMP too big on first fragment */
+		if (!ipvsh->fragoffs)
+			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
+		IP_VS_DBG(1, "frag needed for %pI6c\n", &ipv6_hdr(skb)->saddr);
+		goto err_put;
+	}
+
+	skb_dst_drop(skb);
+	skb_dst_set(skb, &rt->dst);
+
+	return local;
+
+err_put:
+	dst_release(&rt->dst);
+	return -1;
+
+err_unreach:
+	dst_link_failure(skb);
+	return -1;
 }
 #endif
 
@@ -400,32 +490,15 @@ int
 ip_vs_bypass_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		  struct ip_vs_protocol *pp, struct ip_vs_iphdr *ipvsh)
 {
-	struct rtable *rt;			/* Route to the other host */
 	struct iphdr  *iph = ip_hdr(skb);
-	int    mtu;
 
 	EnterFunction(10);
 
-	rt = __ip_vs_get_out_rt(skb, NULL, iph->daddr, IP_VS_RT_MODE_NON_LOCAL,
-				NULL);
-	if (!rt)
-		goto tx_error_icmp;
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if ((skb->len > mtu) && (iph->frag_off & htons(IP_DF)) &&
-	    !skb_is_gso(skb)) {
-		ip_rt_put(rt);
-		icmp_send(skb, ICMP_DEST_UNREACH,ICMP_FRAG_NEEDED, htonl(mtu));
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
+	if (__ip_vs_get_out_rt(skb, NULL, iph->daddr, IP_VS_RT_MODE_NON_LOCAL,
+			       NULL) < 0)
 		goto tx_error;
-	}
 
-	ip_send_check(ip_hdr(skb));
-
-	/* drop old route */
-	skb_dst_drop(skb);
-	skb_dst_set(skb, &rt->dst);
+	ip_send_check(iph);
 
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
@@ -435,8 +508,6 @@ ip_vs_bypass_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	LeaveFunction(10);
 	return NF_STOLEN;
 
- tx_error_icmp:
-	dst_link_failure(skb);
  tx_error:
 	kfree_skb(skb);
 	LeaveFunction(10);
@@ -446,37 +517,13 @@ ip_vs_bypass_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 #ifdef CONFIG_IP_VS_IPV6
 int
 ip_vs_bypass_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
-		     struct ip_vs_protocol *pp, struct ip_vs_iphdr *iph)
+		     struct ip_vs_protocol *pp, struct ip_vs_iphdr *ipvsh)
 {
-	struct rt6_info *rt;			/* Route to the other host */
-	int    mtu;
-
 	EnterFunction(10);
 
-	rt = __ip_vs_get_out_rt_v6(skb, NULL, &iph->daddr.in6, NULL, 0,
-				   IP_VS_RT_MODE_NON_LOCAL);
-	if (!rt)
-		goto tx_error_icmp;
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if (__mtu_check_toobig_v6(skb, mtu)) {
-		if (!skb->dev) {
-			struct net *net = dev_net(skb_dst(skb)->dev);
-
-			skb->dev = net->loopback_dev;
-		}
-		/* only send ICMP too big on first fragment */
-		if (!iph->fragoffs)
-			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
-		dst_release(&rt->dst);
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
+	if (__ip_vs_get_out_rt_v6(skb, NULL, &ipvsh->daddr.in6, NULL,
+				  ipvsh, 0, IP_VS_RT_MODE_NON_LOCAL) < 0)
 		goto tx_error;
-	}
-
-	/* drop old route */
-	skb_dst_drop(skb);
-	skb_dst_set(skb, &rt->dst);
 
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
@@ -486,8 +533,6 @@ ip_vs_bypass_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	LeaveFunction(10);
 	return NF_STOLEN;
 
- tx_error_icmp:
-	dst_link_failure(skb);
  tx_error:
 	kfree_skb(skb);
 	LeaveFunction(10);
@@ -504,28 +549,29 @@ ip_vs_nat_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	       struct ip_vs_protocol *pp, struct ip_vs_iphdr *ipvsh)
 {
 	struct rtable *rt;		/* Route to the other host */
-	int mtu;
-	struct iphdr *iph = ip_hdr(skb);
-	int local, rc;
+	int local, rc, was_input;
 
 	EnterFunction(10);
 
 	/* check if it is a connection of no-client-port */
 	if (unlikely(cp->flags & IP_VS_CONN_F_NO_CPORT)) {
 		__be16 _pt, *p;
-		p = skb_header_pointer(skb, iph->ihl*4, sizeof(_pt), &_pt);
+
+		p = skb_header_pointer(skb, ipvsh->len, sizeof(_pt), &_pt);
 		if (p == NULL)
 			goto tx_error;
 		ip_vs_conn_fill_cport(cp, *p);
 		IP_VS_DBG(10, "filled cport=%d\n", ntohs(*p));
 	}
 
-	if (!(rt = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip,
-				      IP_VS_RT_MODE_LOCAL |
-				      IP_VS_RT_MODE_NON_LOCAL |
-				      IP_VS_RT_MODE_RDR, NULL)))
-		goto tx_error_icmp;
-	local = rt->rt_flags & RTCF_LOCAL;
+	was_input = rt_is_input_route(skb_rtable(skb));
+	local = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip,
+				   IP_VS_RT_MODE_LOCAL |
+				   IP_VS_RT_MODE_NON_LOCAL |
+				   IP_VS_RT_MODE_RDR, NULL);
+	if (local < 0)
+		goto tx_error;
+	rt = skb_rtable(skb);
 	/*
 	 * Avoid duplicate tuple in reply direction for NAT traffic
 	 * to local address when connection is sync-ed
@@ -539,49 +585,31 @@ ip_vs_nat_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 			IP_VS_DBG_RL_PKT(10, AF_INET, pp, skb, 0,
 					 "ip_vs_nat_xmit(): "
 					 "stopping DNAT to local address");
-			goto tx_error_put;
+			goto tx_error;
 		}
 	}
 #endif
 
 	/* From world but DNAT to loopback address? */
-	if (local && ipv4_is_loopback(cp->daddr.ip) &&
-	    rt_is_input_route(skb_rtable(skb))) {
+	if (local && ipv4_is_loopback(cp->daddr.ip) && was_input) {
 		IP_VS_DBG_RL_PKT(1, AF_INET, pp, skb, 0, "ip_vs_nat_xmit(): "
 				 "stopping DNAT to loopback address");
-		goto tx_error_put;
-	}
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if ((skb->len > mtu) && (iph->frag_off & htons(IP_DF)) &&
-	    !skb_is_gso(skb)) {
-		icmp_send(skb, ICMP_DEST_UNREACH,ICMP_FRAG_NEEDED, htonl(mtu));
-		IP_VS_DBG_RL_PKT(0, AF_INET, pp, skb, 0,
-				 "ip_vs_nat_xmit(): frag needed for");
-		goto tx_error_put;
+		goto tx_error;
 	}
 
 	/* copy-on-write the packet before mangling it */
 	if (!skb_make_writable(skb, sizeof(struct iphdr)))
-		goto tx_error_put;
+		goto tx_error;
 
 	if (skb_cow(skb, rt->dst.dev->hard_header_len))
-		goto tx_error_put;
+		goto tx_error;
 
 	/* mangle the packet */
 	if (pp->dnat_handler && !pp->dnat_handler(skb, pp, cp, ipvsh))
-		goto tx_error_put;
+		goto tx_error;
 	ip_hdr(skb)->daddr = cp->daddr.ip;
 	ip_send_check(ip_hdr(skb));
 
-	if (!local) {
-		/* drop old route */
-		skb_dst_drop(skb);
-		skb_dst_set(skb, &rt->dst);
-	} else
-		ip_rt_put(rt);
-
 	IP_VS_DBG_PKT(10, AF_INET, pp, skb, 0, "After DNAT");
 
 	/* FIXME: when application helper enlarges the packet and the length
@@ -596,44 +624,40 @@ ip_vs_nat_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	LeaveFunction(10);
 	return rc;
 
-  tx_error_icmp:
-	dst_link_failure(skb);
   tx_error:
 	kfree_skb(skb);
 	LeaveFunction(10);
 	return NF_STOLEN;
-  tx_error_put:
-	ip_rt_put(rt);
-	goto tx_error;
 }
 
 #ifdef CONFIG_IP_VS_IPV6
 int
 ip_vs_nat_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
-		  struct ip_vs_protocol *pp, struct ip_vs_iphdr *iph)
+		  struct ip_vs_protocol *pp, struct ip_vs_iphdr *ipvsh)
 {
 	struct rt6_info *rt;		/* Route to the other host */
-	int mtu;
 	int local, rc;
 
 	EnterFunction(10);
 
 	/* check if it is a connection of no-client-port */
-	if (unlikely(cp->flags & IP_VS_CONN_F_NO_CPORT && !iph->fragoffs)) {
+	if (unlikely(cp->flags & IP_VS_CONN_F_NO_CPORT && !ipvsh->fragoffs)) {
 		__be16 _pt, *p;
-		p = skb_header_pointer(skb, iph->len, sizeof(_pt), &_pt);
+		p = skb_header_pointer(skb, ipvsh->len, sizeof(_pt), &_pt);
 		if (p == NULL)
 			goto tx_error;
 		ip_vs_conn_fill_cport(cp, *p);
 		IP_VS_DBG(10, "filled cport=%d\n", ntohs(*p));
 	}
 
-	if (!(rt = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6, NULL,
-					 0, (IP_VS_RT_MODE_LOCAL |
-					     IP_VS_RT_MODE_NON_LOCAL |
-					     IP_VS_RT_MODE_RDR))))
-		goto tx_error_icmp;
-	local = __ip_vs_is_local_route6(rt);
+	local = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6, NULL,
+				      ipvsh, 0,
+				      IP_VS_RT_MODE_LOCAL |
+				      IP_VS_RT_MODE_NON_LOCAL |
+				      IP_VS_RT_MODE_RDR);
+	if (local < 0)
+		goto tx_error;
+	rt = (struct rt6_info *) skb_dst(skb);
 	/*
 	 * Avoid duplicate tuple in reply direction for NAT traffic
 	 * to local address when connection is sync-ed
@@ -647,7 +671,7 @@ ip_vs_nat_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 			IP_VS_DBG_RL_PKT(10, AF_INET6, pp, skb, 0,
 					 "ip_vs_nat_xmit_v6(): "
 					 "stopping DNAT to local address");
-			goto tx_error_put;
+			goto tx_error;
 		}
 	}
 #endif
@@ -658,45 +682,21 @@ ip_vs_nat_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 		IP_VS_DBG_RL_PKT(1, AF_INET6, pp, skb, 0,
 				 "ip_vs_nat_xmit_v6(): "
 				 "stopping DNAT to loopback address");
-		goto tx_error_put;
-	}
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if (__mtu_check_toobig_v6(skb, mtu)) {
-		if (!skb->dev) {
-			struct net *net = dev_net(skb_dst(skb)->dev);
-
-			skb->dev = net->loopback_dev;
-		}
-		/* only send ICMP too big on first fragment */
-		if (!iph->fragoffs)
-			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
-		IP_VS_DBG_RL_PKT(0, AF_INET6, pp, skb, 0,
-				 "ip_vs_nat_xmit_v6(): frag needed for");
-		goto tx_error_put;
+		goto tx_error;
 	}
 
 	/* copy-on-write the packet before mangling it */
 	if (!skb_make_writable(skb, sizeof(struct ipv6hdr)))
-		goto tx_error_put;
+		goto tx_error;
 
 	if (skb_cow(skb, rt->dst.dev->hard_header_len))
-		goto tx_error_put;
+		goto tx_error;
 
 	/* mangle the packet */
-	if (pp->dnat_handler && !pp->dnat_handler(skb, pp, cp, iph))
+	if (pp->dnat_handler && !pp->dnat_handler(skb, pp, cp, ipvsh))
 		goto tx_error;
 	ipv6_hdr(skb)->daddr = cp->daddr.in6;
 
-	if (!local || !skb->dev) {
-		skb_dst_drop(skb);
-		skb_dst_set(skb, &rt->dst);
-	} else {
-		/* destined to loopback, do we need to change route? */
-		dst_release(&rt->dst);
-	}
-
 	IP_VS_DBG_PKT(10, AF_INET6, pp, skb, 0, "After DNAT");
 
 	/* FIXME: when application helper enlarges the packet and the length
@@ -711,15 +711,10 @@ ip_vs_nat_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	LeaveFunction(10);
 	return rc;
 
-tx_error_icmp:
-	dst_link_failure(skb);
 tx_error:
 	LeaveFunction(10);
 	kfree_skb(skb);
 	return NF_STOLEN;
-tx_error_put:
-	dst_release(&rt->dst);
-	goto tx_error;
 }
 #endif
 
@@ -756,40 +751,26 @@ ip_vs_tunnel_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	__be16 df;
 	struct iphdr  *iph;			/* Our new IP header */
 	unsigned int max_headroom;		/* The extra header space needed */
-	int    mtu;
-	int ret;
+	int ret, local;
 
 	EnterFunction(10);
 
-	if (!(rt = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip,
-				      IP_VS_RT_MODE_LOCAL |
-				      IP_VS_RT_MODE_NON_LOCAL |
-				      IP_VS_RT_MODE_CONNECT, &saddr)))
-		goto tx_error_icmp;
-	if (rt->rt_flags & RTCF_LOCAL) {
-		ip_rt_put(rt);
+	local = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip,
+				   IP_VS_RT_MODE_LOCAL |
+				   IP_VS_RT_MODE_NON_LOCAL |
+				   IP_VS_RT_MODE_CONNECT |
+				   IP_VS_RT_MODE_TUNNEL, &saddr);
+	if (local < 0)
+		goto tx_error;
+	if (local)
 		return ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 1);
-	}
 
+	rt = skb_rtable(skb);
 	tdev = rt->dst.dev;
 
-	mtu = dst_mtu(&rt->dst) - sizeof(struct iphdr);
-	if (mtu < 68) {
-		IP_VS_DBG_RL("%s(): mtu less than 68\n", __func__);
-		goto tx_error_put;
-	}
-	if (rt_is_output_route(skb_rtable(skb)))
-		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
-
 	/* Copy DF, reset fragment offset and MF */
 	df = sysctl_pmtu_disc(ipvs) ? old_iph->frag_off & htons(IP_DF) : 0;
 
-	if (df && mtu < ntohs(old_iph->tot_len) && !skb_is_gso(skb)) {
-		icmp_send(skb, ICMP_DEST_UNREACH,ICMP_FRAG_NEEDED, htonl(mtu));
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
-		goto tx_error_put;
-	}
-
 	/*
 	 * Okay, now see if we can stuff it in the buffer as-is.
 	 */
@@ -798,12 +779,9 @@ ip_vs_tunnel_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	if (skb_headroom(skb) < max_headroom || skb_cloned(skb)) {
 		struct sk_buff *new_skb =
 			skb_realloc_headroom(skb, max_headroom);
-		if (!new_skb) {
-			ip_rt_put(rt);
-			kfree_skb(skb);
-			IP_VS_ERR_RL("%s(): no memory\n", __func__);
-			return NF_STOLEN;
-		}
+
+		if (!new_skb)
+			goto tx_error;
 		consume_skb(skb);
 		skb = new_skb;
 		old_iph = ip_hdr(skb);
@@ -818,10 +796,6 @@ ip_vs_tunnel_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	skb_reset_network_header(skb);
 	memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
 
-	/* drop old route */
-	skb_dst_drop(skb);
-	skb_dst_set(skb, &rt->dst);
-
 	/*
 	 *	Push down and install the IPIP header.
 	 */
@@ -849,15 +823,10 @@ ip_vs_tunnel_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 
 	return NF_STOLEN;
 
-  tx_error_icmp:
-	dst_link_failure(skb);
   tx_error:
 	kfree_skb(skb);
 	LeaveFunction(10);
 	return NF_STOLEN;
-tx_error_put:
-	ip_rt_put(rt);
-	goto tx_error;
 }
 
 #ifdef CONFIG_IP_VS_IPV6
@@ -871,45 +840,23 @@ ip_vs_tunnel_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	struct ipv6hdr  *old_iph = ipv6_hdr(skb);
 	struct ipv6hdr  *iph;		/* Our new IP header */
 	unsigned int max_headroom;	/* The extra header space needed */
-	int    mtu;
-	int ret;
+	int ret, local;
 
 	EnterFunction(10);
 
-	if (!(rt = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6,
-					 &saddr, 1, (IP_VS_RT_MODE_LOCAL |
-						     IP_VS_RT_MODE_NON_LOCAL))))
-		goto tx_error_icmp;
-	if (__ip_vs_is_local_route6(rt)) {
-		dst_release(&rt->dst);
+	local = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6,
+				      &saddr, ipvsh, 1,
+				      IP_VS_RT_MODE_LOCAL |
+				      IP_VS_RT_MODE_NON_LOCAL |
+				      IP_VS_RT_MODE_TUNNEL);
+	if (local < 0)
+		goto tx_error;
+	if (local)
 		return ip_vs_send_or_cont(NFPROTO_IPV6, skb, cp, 1);
-	}
 
+	rt = (struct rt6_info *) skb_dst(skb);
 	tdev = rt->dst.dev;
 
-	mtu = dst_mtu(&rt->dst) - sizeof(struct ipv6hdr);
-	if (mtu < IPV6_MIN_MTU) {
-		IP_VS_DBG_RL("%s(): mtu less than %d\n", __func__,
-			     IPV6_MIN_MTU);
-		goto tx_error_put;
-	}
-	if (skb_dst(skb))
-		skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
-
-	/* MTU checking: Notice that 'mtu' have been adjusted before hand */
-	if (__mtu_check_toobig_v6(skb, mtu)) {
-		if (!skb->dev) {
-			struct net *net = dev_net(skb_dst(skb)->dev);
-
-			skb->dev = net->loopback_dev;
-		}
-		/* only send ICMP too big on first fragment */
-		if (!ipvsh->fragoffs)
-			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
-		goto tx_error_put;
-	}
-
 	/*
 	 * Okay, now see if we can stuff it in the buffer as-is.
 	 */
@@ -918,12 +865,9 @@ ip_vs_tunnel_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	if (skb_headroom(skb) < max_headroom || skb_cloned(skb)) {
 		struct sk_buff *new_skb =
 			skb_realloc_headroom(skb, max_headroom);
-		if (!new_skb) {
-			dst_release(&rt->dst);
-			kfree_skb(skb);
-			IP_VS_ERR_RL("%s(): no memory\n", __func__);
-			return NF_STOLEN;
-		}
+
+		if (!new_skb)
+			goto tx_error;
 		consume_skb(skb);
 		skb = new_skb;
 		old_iph = ipv6_hdr(skb);
@@ -935,10 +879,6 @@ ip_vs_tunnel_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	skb_reset_network_header(skb);
 	memset(&(IPCB(skb)->opt), 0, sizeof(IPCB(skb)->opt));
 
-	/* drop old route */
-	skb_dst_drop(skb);
-	skb_dst_set(skb, &rt->dst);
-
 	/*
 	 *	Push down and install the IPIP header.
 	 */
@@ -966,15 +906,10 @@ ip_vs_tunnel_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 
 	return NF_STOLEN;
 
-tx_error_icmp:
-	dst_link_failure(skb);
 tx_error:
 	kfree_skb(skb);
 	LeaveFunction(10);
 	return NF_STOLEN;
-tx_error_put:
-	dst_release(&rt->dst);
-	goto tx_error;
 }
 #endif
 
@@ -987,38 +922,21 @@ int
 ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	      struct ip_vs_protocol *pp, struct ip_vs_iphdr *ipvsh)
 {
-	struct rtable *rt;			/* Route to the other host */
-	struct iphdr  *iph = ip_hdr(skb);
-	int    mtu;
+	int local;
 
 	EnterFunction(10);
 
-	if (!(rt = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip,
-				      IP_VS_RT_MODE_LOCAL |
-				      IP_VS_RT_MODE_NON_LOCAL |
-				      IP_VS_RT_MODE_KNOWN_NH, NULL)))
-		goto tx_error_icmp;
-	if (rt->rt_flags & RTCF_LOCAL) {
-		ip_rt_put(rt);
-		return ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 1);
-	}
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if ((iph->frag_off & htons(IP_DF)) && skb->len > mtu &&
-	    !skb_is_gso(skb)) {
-		icmp_send(skb, ICMP_DEST_UNREACH,ICMP_FRAG_NEEDED, htonl(mtu));
-		ip_rt_put(rt);
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
+	local = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip,
+				   IP_VS_RT_MODE_LOCAL |
+				   IP_VS_RT_MODE_NON_LOCAL |
+				   IP_VS_RT_MODE_KNOWN_NH, NULL);
+	if (local < 0)
 		goto tx_error;
-	}
+	if (local)
+		return ip_vs_send_or_cont(NFPROTO_IPV4, skb, cp, 1);
 
 	ip_send_check(ip_hdr(skb));
 
-	/* drop old route */
-	skb_dst_drop(skb);
-	skb_dst_set(skb, &rt->dst);
-
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
@@ -1027,8 +945,6 @@ ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	LeaveFunction(10);
 	return NF_STOLEN;
 
-  tx_error_icmp:
-	dst_link_failure(skb);
   tx_error:
 	kfree_skb(skb);
 	LeaveFunction(10);
@@ -1038,41 +954,20 @@ ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 #ifdef CONFIG_IP_VS_IPV6
 int
 ip_vs_dr_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
-		 struct ip_vs_protocol *pp, struct ip_vs_iphdr *iph)
+		 struct ip_vs_protocol *pp, struct ip_vs_iphdr *ipvsh)
 {
-	struct rt6_info *rt;			/* Route to the other host */
-	int    mtu;
+	int local;
 
 	EnterFunction(10);
 
-	if (!(rt = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6, NULL,
-					 0, (IP_VS_RT_MODE_LOCAL |
-					     IP_VS_RT_MODE_NON_LOCAL))))
-		goto tx_error_icmp;
-	if (__ip_vs_is_local_route6(rt)) {
-		dst_release(&rt->dst);
-		return ip_vs_send_or_cont(NFPROTO_IPV6, skb, cp, 1);
-	}
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if (__mtu_check_toobig_v6(skb, mtu)) {
-		if (!skb->dev) {
-			struct net *net = dev_net(skb_dst(skb)->dev);
-
-			skb->dev = net->loopback_dev;
-		}
-		/* only send ICMP too big on first fragment */
-		if (!iph->fragoffs)
-			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
-		dst_release(&rt->dst);
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
+	local = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6, NULL,
+				      ipvsh, 0,
+				      IP_VS_RT_MODE_LOCAL |
+				      IP_VS_RT_MODE_NON_LOCAL);
+	if (local < 0)
 		goto tx_error;
-	}
-
-	/* drop old route */
-	skb_dst_drop(skb);
-	skb_dst_set(skb, &rt->dst);
+	if (local)
+		return ip_vs_send_or_cont(NFPROTO_IPV6, skb, cp, 1);
 
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
@@ -1082,8 +977,6 @@ ip_vs_dr_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	LeaveFunction(10);
 	return NF_STOLEN;
 
-tx_error_icmp:
-	dst_link_failure(skb);
 tx_error:
 	kfree_skb(skb);
 	LeaveFunction(10);
@@ -1102,10 +995,9 @@ ip_vs_icmp_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 		struct ip_vs_iphdr *iph)
 {
 	struct rtable	*rt;	/* Route to the other host */
-	int mtu;
 	int rc;
 	int local;
-	int rt_mode;
+	int rt_mode, was_input;
 
 	EnterFunction(10);
 
@@ -1125,15 +1017,16 @@ ip_vs_icmp_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 	/*
 	 * mangle and send the packet here (only for VS/NAT)
 	 */
+	was_input = rt_is_input_route(skb_rtable(skb));
 
 	/* LOCALNODE from FORWARD hook is not supported */
 	rt_mode = (hooknum != NF_INET_FORWARD) ?
 		  IP_VS_RT_MODE_LOCAL | IP_VS_RT_MODE_NON_LOCAL |
 		  IP_VS_RT_MODE_RDR : IP_VS_RT_MODE_NON_LOCAL;
-	if (!(rt = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip,
-				      rt_mode, NULL)))
-		goto tx_error_icmp;
-	local = rt->rt_flags & RTCF_LOCAL;
+	local = __ip_vs_get_out_rt(skb, cp->dest, cp->daddr.ip, rt_mode, NULL);
+	if (local < 0)
+		goto tx_error;
+	rt = skb_rtable(skb);
 
 	/*
 	 * Avoid duplicate tuple in reply direction for NAT traffic
@@ -1148,71 +1041,49 @@ ip_vs_icmp_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 			IP_VS_DBG(10, "%s(): "
 				  "stopping DNAT to local address %pI4\n",
 				  __func__, &cp->daddr.ip);
-			goto tx_error_put;
+			goto tx_error;
 		}
 	}
 #endif
 
 	/* From world but DNAT to loopback address? */
-	if (local && ipv4_is_loopback(cp->daddr.ip) &&
-	    rt_is_input_route(skb_rtable(skb))) {
+	if (local && ipv4_is_loopback(cp->daddr.ip) && was_input) {
 		IP_VS_DBG(1, "%s(): "
 			  "stopping DNAT to loopback %pI4\n",
 			  __func__, &cp->daddr.ip);
-		goto tx_error_put;
-	}
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if ((skb->len > mtu) && (ip_hdr(skb)->frag_off & htons(IP_DF)) &&
-	    !skb_is_gso(skb)) {
-		icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
-		goto tx_error_put;
+		goto tx_error;
 	}
 
 	/* copy-on-write the packet before mangling it */
 	if (!skb_make_writable(skb, offset))
-		goto tx_error_put;
+		goto tx_error;
 
 	if (skb_cow(skb, rt->dst.dev->hard_header_len))
-		goto tx_error_put;
+		goto tx_error;
 
 	ip_vs_nat_icmp(skb, pp, cp, 0);
 
-	if (!local) {
-		skb_dst_drop(skb);
-		skb_dst_set(skb, &rt->dst);
-	} else
-		ip_rt_put(rt);
-
 	/* Another hack: avoid icmp_send in ip_fragment */
 	skb->local_df = 1;
 
 	rc = ip_vs_nat_send_or_cont(NFPROTO_IPV4, skb, cp, local);
 	goto out;
 
-  tx_error_icmp:
-	dst_link_failure(skb);
   tx_error:
 	dev_kfree_skb(skb);
 	rc = NF_STOLEN;
   out:
 	LeaveFunction(10);
 	return rc;
-  tx_error_put:
-	ip_rt_put(rt);
-	goto tx_error;
 }
 
 #ifdef CONFIG_IP_VS_IPV6
 int
 ip_vs_icmp_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 		struct ip_vs_protocol *pp, int offset, unsigned int hooknum,
-		struct ip_vs_iphdr *iph)
+		struct ip_vs_iphdr *ipvsh)
 {
 	struct rt6_info	*rt;	/* Route to the other host */
-	int mtu;
 	int rc;
 	int local;
 	int rt_mode;
@@ -1224,7 +1095,7 @@ ip_vs_icmp_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	   translate address/port back */
 	if (IP_VS_FWD_METHOD(cp) != IP_VS_CONN_F_MASQ) {
 		if (cp->packet_xmit)
-			rc = cp->packet_xmit(skb, cp, pp, iph);
+			rc = cp->packet_xmit(skb, cp, pp, ipvsh);
 		else
 			rc = NF_ACCEPT;
 		/* do not touch skb anymore */
@@ -1240,11 +1111,11 @@ ip_vs_icmp_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 	rt_mode = (hooknum != NF_INET_FORWARD) ?
 		  IP_VS_RT_MODE_LOCAL | IP_VS_RT_MODE_NON_LOCAL |
 		  IP_VS_RT_MODE_RDR : IP_VS_RT_MODE_NON_LOCAL;
-	if (!(rt = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6, NULL,
-					 0, rt_mode)))
-		goto tx_error_icmp;
-
-	local = __ip_vs_is_local_route6(rt);
+	local = __ip_vs_get_out_rt_v6(skb, cp->dest, &cp->daddr.in6, NULL,
+				      ipvsh, 0, rt_mode);
+	if (local < 0)
+		goto tx_error;
+	rt = (struct rt6_info *) skb_dst(skb);
 	/*
 	 * Avoid duplicate tuple in reply direction for NAT traffic
 	 * to local address when connection is sync-ed
@@ -1258,7 +1129,7 @@ ip_vs_icmp_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 			IP_VS_DBG(10, "%s(): "
 				  "stopping DNAT to local address %pI6\n",
 				  __func__, &cp->daddr.in6);
-			goto tx_error_put;
+			goto tx_error;
 		}
 	}
 #endif
@@ -1269,57 +1140,29 @@ ip_vs_icmp_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
 		IP_VS_DBG(1, "%s(): "
 			  "stopping DNAT to loopback %pI6\n",
 			  __func__, &cp->daddr.in6);
-		goto tx_error_put;
-	}
-
-	/* MTU checking */
-	mtu = dst_mtu(&rt->dst);
-	if (__mtu_check_toobig_v6(skb, mtu)) {
-		if (!skb->dev) {
-			struct net *net = dev_net(skb_dst(skb)->dev);
-
-			skb->dev = net->loopback_dev;
-		}
-		/* only send ICMP too big on first fragment */
-		if (!iph->fragoffs)
-			icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
-		IP_VS_DBG_RL("%s(): frag needed\n", __func__);
-		goto tx_error_put;
+		goto tx_error;
 	}
 
 	/* copy-on-write the packet before mangling it */
 	if (!skb_make_writable(skb, offset))
-		goto tx_error_put;
+		goto tx_error;
 
 	if (skb_cow(skb, rt->dst.dev->hard_header_len))
-		goto tx_error_put;
+		goto tx_error;
 
 	ip_vs_nat_icmp_v6(skb, pp, cp, 0);
 
-	if (!local || !skb->dev) {
-		skb_dst_drop(skb);
-		skb_dst_set(skb, &rt->dst);
-	} else {
-		/* destined to loopback, do we need to change route? */
-		dst_release(&rt->dst);
-	}
author maintainer
#9 lore raw

PATCH v3 9/15 ipvs: optimize dst usage for real server

In reply to Julian Anastasov

Currently when forwarding requests to real servers
we use dst_lock and atomic operations when cloning the dst_cache value. As the dst_cache value does not change most of the time it is better to use RCU and to lock dst_lock only when we need to replace the obsoleted dst. For this to work we keep dst_cache in new structure protected by RCU. For packets to remote real servers we will use noref version of dst_cache, it will be valid while we are in RCU read-side critical section because now dst_release for replaced dsts will be invoked after the grace period. Packets to local real servers that are passed to local stack with NF_ACCEPT need a dst clone.

---
 include/net/ip_vs.h             |   12 ++-
 net/netfilter/ipvs/ip_vs_core.c |   11 ++-
 net/netfilter/ipvs/ip_vs_ctl.c  |   25 ++++--
 net/netfilter/ipvs/ip_vs_xmit.c |  188 +++++++++++++++++++++++++++++---------
 4 files changed, 177 insertions(+), 59 deletions(-)
include/net/ip_vs.h +9 -3
index c448d3c..c0185dd 100644
@@ -724,6 +724,13 @@ struct ip_vs_service {
 	struct ip_vs_pe		*pe;
 };
 
+/* Information for cached dst */
+struct ip_vs_dest_dst {
+	struct dst_entry	*dst_cache;	/* destination cache entry */
+	u32			dst_cookie;
+	union nf_inet_addr	dst_saddr;
+	struct rcu_head		rcu_head;
+};
 
 /*
  *	The real server destination forwarding entry
@@ -752,9 +759,7 @@ struct ip_vs_dest {
 
 	/* for destination cache */
 	spinlock_t		dst_lock;	/* lock of dst_cache */
-	struct dst_entry	*dst_cache;	/* destination cache entry */
-	u32			dst_cookie;
-	union nf_inet_addr	dst_saddr;
+	struct ip_vs_dest_dst __rcu *dest_dst;	/* cached dst info */
 
 	/* for virtual service */
 	struct ip_vs_service	*svc;		/* service it belongs to */
@@ -1427,6 +1432,7 @@ extern int ip_vs_dr_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 extern int ip_vs_icmp_xmit(struct sk_buff *skb, struct ip_vs_conn *cp,
 			   struct ip_vs_protocol *pp, int offset,
 			   unsigned int hooknum, struct ip_vs_iphdr *iph);
+extern void ip_vs_dest_dst_rcu_free(struct rcu_head *head);
 
 #ifdef CONFIG_IP_VS_IPV6
 extern int ip_vs_bypass_xmit_v6(struct sk_buff *skb, struct ip_vs_conn *cp,
net/netfilter/ipvs/ip_vs_core.c +7 -4
index 61f49d2..b97d0c7 100644
@@ -1401,10 +1401,13 @@ ip_vs_in_icmp(struct sk_buff *skb, int *related, unsigned int hooknum)
 				goto ignore_ipip;
 			/* Prefer the resulting PMTU */
 			if (dest) {
-				spin_lock(&dest->dst_lock);
-				if (dest->dst_cache)
-					mtu = dst_mtu(dest->dst_cache);
-				spin_unlock(&dest->dst_lock);
+				struct ip_vs_dest_dst *dest_dst;
+
+				rcu_read_lock();
+				dest_dst = rcu_dereference(dest->dest_dst);
+				if (dest_dst)
+					mtu = dst_mtu(dest_dst->dst_cache);
+				rcu_read_unlock();
 			}
 			if (mtu > 68 + sizeof(struct iphdr))
 				mtu -= sizeof(struct iphdr);
net/netfilter/ipvs/ip_vs_ctl.c +18 -7
index b545549..811cba2 100644
@@ -639,15 +639,26 @@ struct ip_vs_dest *ip_vs_find_dest(struct net  *net, int af,
 	return dest;
 }
 
-/* Release dst_cache for dest in user context */
+void ip_vs_dest_dst_rcu_free(struct rcu_head *head)
+{
+	struct ip_vs_dest_dst *dest_dst = container_of(head,
+						       struct ip_vs_dest_dst,
+						       rcu_head);
+
+	dst_release(dest_dst->dst_cache);
+	kfree(dest_dst);
+}
+
+/* Release dest_dst and dst_cache for dest in user context */
 static void __ip_vs_dst_cache_reset(struct ip_vs_dest *dest)
 {
-	struct dst_entry *old_dst;
+	struct ip_vs_dest_dst *old;
 
-	old_dst = dest->dst_cache;
-	dest->dst_cache = NULL;
-	dst_release(old_dst);
-	dest->dst_saddr.ip = 0;
+	old = rcu_dereference_protected(dest->dest_dst, 1);
+	if (old) {
+		RCU_INIT_POINTER(dest->dest_dst, NULL);
+		call_rcu(&old->rcu_head, ip_vs_dest_dst_rcu_free);
+	}
 }
 
 /*
@@ -1511,7 +1522,7 @@ static inline void
 ip_vs_forget_dev(struct ip_vs_dest *dest, struct net_device *dev)
 {
 	spin_lock_bh(&dest->dst_lock);
-	if (dest->dst_cache && dest->dst_cache->dev == dev) {
+	if (dest->dest_dst && dest->dest_dst->dst_cache->dev == dev) {
 		IP_VS_DBG_BUF(3, "Reset dev:%s dest %s:%u ,dest->refcnt=%d\n",
 			      dev->name,
 			      IP_VS_DBG_ADDR(dest->af, &dest->addr),
net/netfilter/ipvs/ip_vs_xmit.c +143 -45

499 lines not on this page · raw message

Signature · 1 line
1.7.3.4
author maintainer
#10 lore raw

PATCH v3 10/15 ipvs: convert app locks

In reply to Julian Anastasov

We use locks like tcp_app_lock, udp_app_lock,
sctp_app_lock to protect access to the protocol hash tables from readers in packet context while the application instances (inc) are [un]registered under global mutex.

As the hash tables are mostly read when conns are
created and bound to app, use RCU for readers and reclaim app instance after grace period.

Simplify ip_vs_app_inc_get because we use usecnt
only for statistics and rely on module refcounting.

---
 include/net/ip_vs.h                   |    4 +---
 net/netfilter/ipvs/ip_vs_app.c        |   27 +++++++++++++++++++--------
 net/netfilter/ipvs/ip_vs_ftp.c        |    2 ++
 net/netfilter/ipvs/ip_vs_proto_sctp.c |   18 ++++++------------
 net/netfilter/ipvs/ip_vs_proto_tcp.c  |   18 ++++++------------
 net/netfilter/ipvs/ip_vs_proto_udp.c  |   19 ++++++-------------
 6 files changed, 40 insertions(+), 48 deletions(-)
include/net/ip_vs.h +1 -3
index c0185dd..564b30d 100644
@@ -823,6 +823,7 @@ struct ip_vs_app {
 	struct ip_vs_app	*app;		/* its real application */
 	__be16			port;		/* port number in net order */
 	atomic_t		usecnt;		/* usage counter */
+	struct rcu_head		rcu_head;
 
 	/*
 	 * output hook: Process packet in inout direction, diff set for TCP.
@@ -908,7 +909,6 @@ struct netns_ipvs {
 	#define	TCP_APP_TAB_SIZE	(1 << TCP_APP_TAB_BITS)
 	#define	TCP_APP_TAB_MASK	(TCP_APP_TAB_SIZE - 1)
 	struct list_head	tcp_apps[TCP_APP_TAB_SIZE];
-	spinlock_t		tcp_app_lock;
 #endif
 	/* ip_vs_proto_udp */
 #ifdef CONFIG_IP_VS_PROTO_UDP
@@ -916,7 +916,6 @@ struct netns_ipvs {
 	#define	UDP_APP_TAB_SIZE	(1 << UDP_APP_TAB_BITS)
 	#define	UDP_APP_TAB_MASK	(UDP_APP_TAB_SIZE - 1)
 	struct list_head	udp_apps[UDP_APP_TAB_SIZE];
-	spinlock_t		udp_app_lock;
 #endif
 	/* ip_vs_proto_sctp */
 #ifdef CONFIG_IP_VS_PROTO_SCTP
@@ -925,7 +924,6 @@ struct netns_ipvs {
 	#define SCTP_APP_TAB_MASK	(SCTP_APP_TAB_SIZE - 1)
 	/* Hash table for SCTP application incarnations	 */
 	struct list_head	sctp_apps[SCTP_APP_TAB_SIZE];
-	spinlock_t		sctp_app_lock;
 #endif
 	/* ip_vs_conn */
 	atomic_t		conn_count;      /*  connection counter */
net/netfilter/ipvs/ip_vs_app.c +19 -8
index 0b779d7..a956030 100644
@@ -58,6 +58,18 @@ static inline void ip_vs_app_put(struct ip_vs_app *app)
 	module_put(app->module);
 }
 
+static void ip_vs_app_inc_destroy(struct ip_vs_app *inc)
+{
+	kfree(inc->timeout_table);
+	kfree(inc);
+}
+
+static void ip_vs_app_inc_rcu_free(struct rcu_head *head)
+{
+	struct ip_vs_app *inc = container_of(head, struct ip_vs_app, rcu_head);
+
+	ip_vs_app_inc_destroy(inc);
+}
 
 /*
  *	Allocate/initialize app incarnation and register it in proto apps.
@@ -106,8 +118,7 @@ ip_vs_app_inc_new(struct net *net, struct ip_vs_app *app, __u16 proto,
 	return 0;
 
   out:
-	kfree(inc->timeout_table);
-	kfree(inc);
+	ip_vs_app_inc_destroy(inc);
 	return ret;
 }
 
@@ -131,8 +142,7 @@ ip_vs_app_inc_release(struct net *net, struct ip_vs_app *inc)
 
 	list_del(&inc->a_list);
 
-	kfree(inc->timeout_table);
-	kfree(inc);
+	call_rcu(&inc->rcu_head, ip_vs_app_inc_rcu_free);
 }
 
 
@@ -144,9 +154,9 @@ int ip_vs_app_inc_get(struct ip_vs_app *inc)
 {
 	int result;
 
-	atomic_inc(&inc->usecnt);
-	if (unlikely((result = ip_vs_app_get(inc->app)) != 1))
-		atomic_dec(&inc->usecnt);
+	result = ip_vs_app_get(inc->app);
+	if (result)
+		atomic_inc(&inc->usecnt);
 	return result;
 }
 
@@ -156,8 +166,8 @@ int ip_vs_app_inc_get(struct ip_vs_app *inc)
  */
 void ip_vs_app_inc_put(struct ip_vs_app *inc)
 {
-	ip_vs_app_put(inc->app);
 	atomic_dec(&inc->usecnt);
+	ip_vs_app_put(inc->app);
 }
 
 
@@ -218,6 +228,7 @@ out_unlock:
 /*
  *	ip_vs_app unregistration routine
  *	We are sure there are no app incarnations attached to services
+ *	Caller should use synchronize_rcu() or rcu_barrier()
  */
 void unregister_ip_vs_app(struct net *net, struct ip_vs_app *app)
 {
net/netfilter/ipvs/ip_vs_ftp.c +2 -0
index 4f53a5f..7f90825 100644
@@ -480,6 +480,7 @@ static int __init ip_vs_ftp_init(void)
 	int rv;
 
 	rv = register_pernet_subsys(&ip_vs_ftp_ops);
+	/* rcu_barrier() is called by netns on error */
 	return rv;
 }
 
@@ -489,6 +490,7 @@ static int __init ip_vs_ftp_init(void)
 static void __exit ip_vs_ftp_exit(void)
 {
 	unregister_pernet_subsys(&ip_vs_ftp_ops);
+	/* rcu_barrier() is called by netns */
 }
 
 
net/netfilter/ipvs/ip_vs_proto_sctp.c +6 -12
index cd1d729..f7190cd 100644
@@ -1016,30 +1016,25 @@ static int sctp_register_app(struct net *net, struct ip_vs_app *inc)
 
 	hash = sctp_app_hashkey(port);
 
-	spin_lock_bh(&ipvs->sctp_app_lock);
 	list_for_each_entry(i, &ipvs->sctp_apps[hash], p_list) {
 		if (i->port == port) {
 			ret = -EEXIST;
 			goto out;
 		}
 	}
-	list_add(&inc->p_list, &ipvs->sctp_apps[hash]);
+	list_add_rcu(&inc->p_list, &ipvs->sctp_apps[hash]);
 	atomic_inc(&pd->appcnt);
 out:
-	spin_unlock_bh(&ipvs->sctp_app_lock);
 
 	return ret;
 }
 
 static void sctp_unregister_app(struct net *net, struct ip_vs_app *inc)
 {
-	struct netns_ipvs *ipvs = net_ipvs(net);
 	struct ip_vs_proto_data *pd = ip_vs_proto_data_get(net, IPPROTO_SCTP);
 
-	spin_lock_bh(&ipvs->sctp_app_lock);
 	atomic_dec(&pd->appcnt);
-	list_del(&inc->p_list);
-	spin_unlock_bh(&ipvs->sctp_app_lock);
+	list_del_rcu(&inc->p_list);
 }
 
 static int sctp_app_conn_bind(struct ip_vs_conn *cp)
@@ -1055,12 +1050,12 @@ static int sctp_app_conn_bind(struct ip_vs_conn *cp)
 	/* Lookup application incarnations and bind the right one */
 	hash = sctp_app_hashkey(cp->vport);
 
-	spin_lock(&ipvs->sctp_app_lock);
-	list_for_each_entry(inc, &ipvs->sctp_apps[hash], p_list) {
+	rcu_read_lock();
+	list_for_each_entry_rcu(inc, &ipvs->sctp_apps[hash], p_list) {
 		if (inc->port == cp->vport) {
 			if (unlikely(!ip_vs_app_inc_get(inc)))
 				break;
-			spin_unlock(&ipvs->sctp_app_lock);
+			rcu_read_unlock();
 
 			IP_VS_DBG_BUF(9, "%s: Binding conn %s:%u->"
 					"%s:%u to app %s on port %u\n",
@@ -1076,7 +1071,7 @@ static int sctp_app_conn_bind(struct ip_vs_conn *cp)
 			goto out;
 		}
 	}
-	spin_unlock(&ipvs->sctp_app_lock);
+	rcu_read_unlock();
 out:
 	return result;
 }
@@ -1090,7 +1085,6 @@ static int __ip_vs_sctp_init(struct net *net, struct ip_vs_proto_data *pd)
 	struct netns_ipvs *ipvs = net_ipvs(net);
 
 	ip_vs_init_hash_table(ipvs->sctp_apps, SCTP_APP_TAB_SIZE);
-	spin_lock_init(&ipvs->sctp_app_lock);
 	pd->timeout_table = ip_vs_create_timeout_table((int *)sctp_timeouts,
 							sizeof(sctp_timeouts));
 	if (!pd->timeout_table)
net/netfilter/ipvs/ip_vs_proto_tcp.c +6 -12
index 9af653a..0bbc3fe 100644
@@ -580,18 +580,16 @@ static int tcp_register_app(struct net *net, struct ip_vs_app *inc)
 
 	hash = tcp_app_hashkey(port);
 
-	spin_lock_bh(&ipvs->tcp_app_lock);
 	list_for_each_entry(i, &ipvs->tcp_apps[hash], p_list) {
 		if (i->port == port) {
 			ret = -EEXIST;
 			goto out;
 		}
 	}
-	list_add(&inc->p_list, &ipvs->tcp_apps[hash]);
+	list_add_rcu(&inc->p_list, &ipvs->tcp_apps[hash]);
 	atomic_inc(&pd->appcnt);
 
   out:
-	spin_unlock_bh(&ipvs->tcp_app_lock);
 	return ret;
 }
 
@@ -599,13 +597,10 @@ static int tcp_register_app(struct net *net, struct ip_vs_app *inc)
 static void
 tcp_unregister_app(struct net *net, struct ip_vs_app *inc)
 {
-	struct netns_ipvs *ipvs = net_ipvs(net);
 	struct ip_vs_proto_data *pd = ip_vs_proto_data_get(net, IPPROTO_TCP);
 
-	spin_lock_bh(&ipvs->tcp_app_lock);
 	atomic_dec(&pd->appcnt);
-	list_del(&inc->p_list);
-	spin_unlock_bh(&ipvs->tcp_app_lock);
+	list_del_rcu(&inc->p_list);
 }
 
 
@@ -624,12 +619,12 @@ tcp_app_conn_bind(struct ip_vs_conn *cp)
 	/* Lookup application incarnations and bind the right one */
 	hash = tcp_app_hashkey(cp->vport);
 
-	spin_lock(&ipvs->tcp_app_lock);
-	list_for_each_entry(inc, &ipvs->tcp_apps[hash], p_list) {
+	rcu_read_lock();
+	list_for_each_entry_rcu(inc, &ipvs->tcp_apps[hash], p_list) {
 		if (inc->port == cp->vport) {
 			if (unlikely(!ip_vs_app_inc_get(inc)))
 				break;
-			spin_unlock(&ipvs->tcp_app_lock);
+			rcu_read_unlock();
 
 			IP_VS_DBG_BUF(9, "%s(): Binding conn %s:%u->"
 				      "%s:%u to app %s on port %u\n",
@@ -646,7 +641,7 @@ tcp_app_conn_bind(struct ip_vs_conn *cp)
 			goto out;
 		}
 	}
-	spin_unlock(&ipvs->tcp_app_lock);
+	rcu_read_unlock();
 
   out:
 	return result;
@@ -676,7 +671,6 @@ static int __ip_vs_tcp_init(struct net *net, struct ip_vs_proto_data *pd)
 	struct netns_ipvs *ipvs = net_ipvs(net);
 
 	ip_vs_init_hash_table(ipvs->tcp_apps, TCP_APP_TAB_SIZE);
-	spin_lock_init(&ipvs->tcp_app_lock);
 	pd->timeout_table = ip_vs_create_timeout_table((int *)tcp_timeouts,
 							sizeof(tcp_timeouts));
 	if (!pd->timeout_table)
net/netfilter/ipvs/ip_vs_proto_udp.c +6 -13
index 503a842..1a03e2d 100644
@@ -359,19 +359,16 @@ static int udp_register_app(struct net *net, struct ip_vs_app *inc)
 
 	hash = udp_app_hashkey(port);
 
-
-	spin_lock_bh(&ipvs->udp_app_lock);
 	list_for_each_entry(i, &ipvs->udp_apps[hash], p_list) {
 		if (i->port == port) {
 			ret = -EEXIST;
 			goto out;
 		}
 	}
-	list_add(&inc->p_list, &ipvs->udp_apps[hash]);
+	list_add_rcu(&inc->p_list, &ipvs->udp_apps[hash]);
 	atomic_inc(&pd->appcnt);
 
   out:
-	spin_unlock_bh(&ipvs->udp_app_lock);
 	return ret;
 }
 
@@ -380,12 +377,9 @@ static void
 udp_unregister_app(struct net *net, struct ip_vs_app *inc)
 {
 	struct ip_vs_proto_data *pd = ip_vs_proto_data_get(net, IPPROTO_UDP);
-	struct netns_ipvs *ipvs = net_ipvs(net);
 
-	spin_lock_bh(&ipvs->udp_app_lock);
 	atomic_dec(&pd->appcnt);
-	list_del(&inc->p_list);
-	spin_unlock_bh(&ipvs->udp_app_lock);
+	list_del_rcu(&inc->p_list);
 }
 
 
@@ -403,12 +397,12 @@ static int udp_app_conn_bind(struct ip_vs_conn *cp)
 	/* Lookup application incarnations and bind the right one */
 	hash = udp_app_hashkey(cp->vport);
 
-	spin_lock(&ipvs->udp_app_lock);
-	list_for_each_entry(inc, &ipvs->udp_apps[hash], p_list) {
+	rcu_read_lock();
+	list_for_each_entry_rcu(inc, &ipvs->udp_apps[hash], p_list) {
 		if (inc->port == cp->vport) {
 			if (unlikely(!ip_vs_app_inc_get(inc)))
 				break;
-			spin_unlock(&ipvs->udp_app_lock);
+			rcu_read_unlock();
 
 			IP_VS_DBG_BUF(9, "%s(): Binding conn %s:%u->"
 				      "%s:%u to app %s on port %u\n",
@@ -425,7 +419,7 @@ static int udp_app_conn_bind(struct ip_vs_conn *cp)
 			goto out;
 		}
 	}
-	spin_unlock(&ipvs->udp_app_lock);
+	rcu_read_unlock();
 
   out:
 	return result;
@@ -467,7 +461,6 @@ static int __udp_init(struct net *net, struct ip_vs_proto_data *pd)
 	struct netns_ipvs *ipvs = net_ipvs(net);
 
 	ip_vs_init_hash_table(ipvs->udp_apps, UDP_APP_TAB_SIZE);
-	spin_lock_init(&ipvs->udp_app_lock);
 	pd->timeout_table = ip_vs_create_timeout_table((int *)udp_timeouts,
 							sizeof(udp_timeouts));
 	if (!pd->timeout_table)
Signature · 1 line
1.7.3.4
author maintainer
#11 lore raw

PATCH v3 11/15 ipvs: remove rs_lock by using RCU

In reply to Julian Anastasov

rs_lock was used to protect rs_table (hash table)
from updaters (under global mutex) and readers (packet handlers). We can remove rs_lock by using RCU lock for readers. Reclaiming dest only with kfree_rcu is enough because the readers access only fields from the ip_vs_dest structure.

Use hlist for rs_table.

As we are now using hlist_del_rcu, introduce in_rs_table
flag as replacement for the list_empty checks which do not work with RCU. It is needed because only NAT dests are in the rs_table.

---
 include/net/ip_vs.h             |   14 ++++---
 net/netfilter/ipvs/ip_vs_core.c |    5 +-
 net/netfilter/ipvs/ip_vs_ctl.c  |   88 ++++++++++++++-------------------------
 3 files changed, 41 insertions(+), 66 deletions(-)
include/net/ip_vs.h +8 -6
index 564b30d..6873d2d 100644
@@ -738,7 +738,7 @@ struct ip_vs_dest_dst {
  */
 struct ip_vs_dest {
 	struct list_head	n_list;   /* for the dests in the service */
-	struct list_head	d_list;   /* for table with all the dests */
+	struct hlist_node	d_list;   /* for table with all the dests */
 
 	u16			af;		/* address family */
 	__be16			port;		/* port number of the server */
@@ -767,6 +767,9 @@ struct ip_vs_dest {
 	__be16			vport;		/* virtual port number */
 	union nf_inet_addr	vaddr;		/* virtual IP address */
 	__u32			vfwmark;	/* firewall mark of service */
+
+	struct rcu_head		rcu_head;
+	unsigned int		in_rs_table:1;	/* we are in rs_table */
 };
 
 
@@ -897,7 +900,7 @@ struct netns_ipvs {
 	#define IP_VS_RTAB_SIZE (1 << IP_VS_RTAB_BITS)
 	#define IP_VS_RTAB_MASK (IP_VS_RTAB_SIZE - 1)
 
-	struct list_head	rs_table[IP_VS_RTAB_SIZE];
+	struct hlist_head	rs_table[IP_VS_RTAB_SIZE];
 	/* ip_vs_app */
 	struct list_head	app_list;
 	/* ip_vs_proto */
@@ -933,7 +936,6 @@ struct netns_ipvs {
 
 	int			num_services;    /* no of virtual services */
 
-	rwlock_t		rs_lock;         /* real services table */
 	/* Trash for destinations */
 	struct list_head	dest_trash;
 	/* Service counters */
@@ -1376,9 +1378,9 @@ static inline void ip_vs_service_put(struct ip_vs_service *svc)
 	atomic_dec(&svc->usecnt);
 }
 
-extern struct ip_vs_dest *
-ip_vs_lookup_real_service(struct net *net, int af, __u16 protocol,
-			  const union nf_inet_addr *daddr, __be16 dport);
+extern bool
+ip_vs_has_real_service(struct net *net, int af, __u16 protocol,
+		       const union nf_inet_addr *daddr, __be16 dport);
 
 extern int ip_vs_use_count_inc(void);
 extern void ip_vs_use_count_dec(void);
net/netfilter/ipvs/ip_vs_core.c +2 -3
index b97d0c7..207c3b9 100644
@@ -1164,9 +1164,8 @@ ip_vs_out(unsigned int hooknum, struct sk_buff *skb, int af)
 					 sizeof(_ports), _ports, &iph);
 		if (pptr == NULL)
 			return NF_ACCEPT;	/* Not for me */
-		if (ip_vs_lookup_real_service(net, af, iph.protocol,
-					      &iph.saddr,
-					      pptr[0])) {
+		if (ip_vs_has_real_service(net, af, iph.protocol, &iph.saddr,
+					   pptr[0])) {
 			/*
 			 * Notify the real server: there is no
 			 * existing entry if it is not RST
net/netfilter/ipvs/ip_vs_ctl.c +29 -54

165 lines not on this page · raw message

author maintainer
#12 lore raw

PATCH v3 12/15 ipvs: convert locks used in persistence engines

In reply to Julian Anastasov

Allow the readers to use RCU lock and for
PE module registrations use global mutex instead of spinlock. All PE modules need to use synchronize_rcu in their module exit handler.

---
 net/netfilter/ipvs/ip_vs_pe.c     |   43 +++++++++++-------------------------
 net/netfilter/ipvs/ip_vs_pe_sip.c |    1 +
 2 files changed, 14 insertions(+), 30 deletions(-)
net/netfilter/ipvs/ip_vs_pe.c +10 -27
index 5cf859c..5d9774c 100644
@@ -13,8 +13,8 @@
 /* IPVS pe list */
 static LIST_HEAD(ip_vs_pe);
 
-/* lock for service table */
-static DEFINE_SPINLOCK(ip_vs_pe_lock);
+/* semaphore for IPVS PEs. */
+static DEFINE_MUTEX(ip_vs_pe_mutex);
 
 /* Bind a service with a pe */
 void ip_vs_bind_pe(struct ip_vs_service *svc, struct ip_vs_pe *pe)
@@ -36,9 +36,8 @@ struct ip_vs_pe *__ip_vs_pe_getbyname(const char *pe_name)
 	IP_VS_DBG(10, "%s(): pe_name \"%s\"\n", __func__,
 		  pe_name);
 
-	spin_lock_bh(&ip_vs_pe_lock);
-
-	list_for_each_entry(pe, &ip_vs_pe, n_list) {
+	rcu_read_lock();
+	list_for_each_entry_rcu(pe, &ip_vs_pe, n_list) {
 		/* Test and get the modules atomically */
 		if (pe->module &&
 		    !try_module_get(pe->module)) {
@@ -47,14 +46,14 @@ struct ip_vs_pe *__ip_vs_pe_getbyname(const char *pe_name)
 		}
 		if (strcmp(pe_name, pe->name)==0) {
 			/* HIT */
-			spin_unlock_bh(&ip_vs_pe_lock);
+			rcu_read_unlock();
 			return pe;
 		}
 		if (pe->module)
 			module_put(pe->module);
 	}
+	rcu_read_unlock();
 
-	spin_unlock_bh(&ip_vs_pe_lock);
 	return NULL;
 }
 
@@ -83,22 +82,13 @@ int register_ip_vs_pe(struct ip_vs_pe *pe)
 	/* increase the module use count */
 	ip_vs_use_count_inc();
 
-	spin_lock_bh(&ip_vs_pe_lock);
-
-	if (!list_empty(&pe->n_list)) {
-		spin_unlock_bh(&ip_vs_pe_lock);
-		ip_vs_use_count_dec();
-		pr_err("%s(): [%s] pe already linked\n",
-		       __func__, pe->name);
-		return -EINVAL;
-	}
-
+	mutex_lock(&ip_vs_pe_mutex);
 	/* Make sure that the pe with this name doesn't exist
 	 * in the pe list.
 	 */
 	list_for_each_entry(tmp, &ip_vs_pe, n_list) {
 		if (strcmp(tmp->name, pe->name) == 0) {
-			spin_unlock_bh(&ip_vs_pe_lock);
+			mutex_unlock(&ip_vs_pe_mutex);
 			ip_vs_use_count_dec();
 			pr_err("%s(): [%s] pe already existed "
 			       "in the system\n", __func__, pe->name);
@@ -106,8 +96,8 @@ int register_ip_vs_pe(struct ip_vs_pe *pe)
 		}
 	}
 	/* Add it into the d-linked pe list */
-	list_add(&pe->n_list, &ip_vs_pe);
-	spin_unlock_bh(&ip_vs_pe_lock);
+	list_add_rcu(&pe->n_list, &ip_vs_pe);
+	mutex_unlock(&ip_vs_pe_mutex);
 
 	pr_info("[%s] pe registered.\n", pe->name);
 
@@ -118,17 +108,10 @@ EXPORT_SYMBOL_GPL(register_ip_vs_pe);
 /* Unregister a pe from the pe list */
 int unregister_ip_vs_pe(struct ip_vs_pe *pe)
 {
-	spin_lock_bh(&ip_vs_pe_lock);
-	if (list_empty(&pe->n_list)) {
-		spin_unlock_bh(&ip_vs_pe_lock);
-		pr_err("%s(): [%s] pe is not in the list. failed\n",
-		       __func__, pe->name);
-		return -EINVAL;
-	}
author maintainer
#13 lore raw

PATCH v3 13/15 ipvs: convert connection locking

In reply to Julian Anastasov

Convert __ip_vs_conntbl_lock_array as follows:

- readers that do not modify conn lists will use RCU lock
- updaters that modify lists will use spinlock_t

Now for conn lookups we will use RCU read-side
critical section. Without using __ip_vs_conn_get such places have access to connection fields and can dereference some pointers like pe and pe_data plus the ability to update timer expiration. If full access is required we contend for reference.

We add barrier in __ip_vs_conn_put, so that
other CPUs see the refcnt operation after other writes.

With the introduction of ip_vs_conn_unlink()
we try to reorganize ip_vs_conn_expire(), so that unhashing of connections that should stay more time is avoided, even if it is for very short time.

---
 include/net/ip_vs.h             |   12 ++
 net/netfilter/ipvs/ip_vs_conn.c |  230 +++++++++++++++++++++------------------
 2 files changed, 134 insertions(+), 108 deletions(-)
include/net/ip_vs.h +12 -0
index 6873d2d..87fadef 100644
@@ -620,6 +620,8 @@ struct ip_vs_conn {
 	const struct ip_vs_pe	*pe;
 	char			*pe_data;
 	__u8			pe_data_len;
+
+	struct rcu_head		rcu_head;
 };
 
 /*
@@ -1185,9 +1187,19 @@ struct ip_vs_conn * ip_vs_conn_out_get_proto(int af, const struct sk_buff *skb,
 					     const struct ip_vs_iphdr *iph,
 					     int inverse);
 
+/* Get reference to gain full access to conn.
+ * By default, RCU read-side critical sections have access only to
+ * conn fields and its PE data, see ip_vs_conn_rcu_free() for reference.
+ */
+static inline bool __ip_vs_conn_get(struct ip_vs_conn *cp)
+{
+	return atomic_inc_not_zero(&cp->refcnt);
+}
+
 /* put back the conn without restarting its timer */
 static inline void __ip_vs_conn_put(struct ip_vs_conn *cp)
 {
+	smp_mb__before_atomic_dec();
 	atomic_dec(&cp->refcnt);
 }
 extern void ip_vs_conn_put(struct ip_vs_conn *cp);
net/netfilter/ipvs/ip_vs_conn.c +122 -108

469 lines not on this page · raw message

Signature · 1 line
1.7.3.4
author maintainer
#14 lore raw

PATCH v3 14/15 ipvs: reorder keys in connection structure

In reply to Julian Anastasov

__ip_vs_conn_in_get and ip_vs_conn_out_get are
hot places. Optimize them, so that ports are matched first. By moving net and fwmark below, on 32-bit arch we can fit caddr in 32-byte cache line and all addresses in 64-byte cache line.

---
 include/net/ip_vs.h             |   12 ++++++------
 net/netfilter/ipvs/ip_vs_conn.c |   19 ++++++++++---------
 2 files changed, 16 insertions(+), 15 deletions(-)
include/net/ip_vs.h +6 -6
index 87fadef..26f104a 100644
@@ -566,20 +566,19 @@ struct ip_vs_conn_param {
  */
 struct ip_vs_conn {
 	struct hlist_node	c_list;         /* hashed list heads */
-#ifdef CONFIG_NET_NS
-	struct net              *net;           /* Name space */
-#endif
 	/* Protocol, addresses and port numbers */
-	u16                     af;             /* address family */
 	__be16                  cport;
-	__be16                  vport;
 	__be16                  dport;
-	__u32                   fwmark;         /* Fire wall mark from skb */
+	__be16                  vport;
+	u16			af;		/* address family */
 	union nf_inet_addr      caddr;          /* client address */
 	union nf_inet_addr      vaddr;          /* virtual address */
 	union nf_inet_addr      daddr;          /* destination address */
 	volatile __u32          flags;          /* status flags */
 	__u16                   protocol;       /* Which protocol (TCP/UDP) */
+#ifdef CONFIG_NET_NS
+	struct net              *net;           /* Name space */
+#endif
 
 	/* counter and timer */
 	atomic_t		refcnt;		/* reference count */
@@ -593,6 +592,7 @@ struct ip_vs_conn {
 						 * state transition triggerd
 						 * synchronization
 						 */
+	__u32			fwmark;		/* Fire wall mark from skb */
 	unsigned long		sync_endtime;	/* jiffies + sent_retries */
 
 	/* Control members */
net/netfilter/ipvs/ip_vs_conn.c +10 -9
index b0cd2be..416015b 100644
@@ -265,8 +265,8 @@ __ip_vs_conn_in_get(const struct ip_vs_conn_param *p)
 	rcu_read_lock();
 
 	hlist_for_each_entry_rcu(cp, &ip_vs_conn_tab[hash], c_list) {
-		if (cp->af == p->af &&
-		    p->cport == cp->cport && p->vport == cp->vport &&
+		if (p->cport == cp->cport && p->vport == cp->vport &&
+		    cp->af == p->af &&
 		    ip_vs_addr_equal(p->af, p->caddr, &cp->caddr) &&
 		    ip_vs_addr_equal(p->af, p->vaddr, &cp->vaddr) &&
 		    ((!p->cport) ^ (!(cp->flags & IP_VS_CONN_F_NO_CPORT))) &&
@@ -350,9 +350,9 @@ struct ip_vs_conn *ip_vs_ct_in_get(const struct ip_vs_conn_param *p)
 	rcu_read_lock();
 
 	hlist_for_each_entry_rcu(cp, &ip_vs_conn_tab[hash], c_list) {
-		if (!ip_vs_conn_net_eq(cp, p->net))
-			continue;
-		if (p->pe_data && p->pe->ct_match) {
+		if (unlikely(p->pe_data && p->pe->ct_match)) {
+			if (!ip_vs_conn_net_eq(cp, p->net))
+				continue;
 			if (p->pe == cp->pe && p->pe->ct_match(p, cp)) {
 				if (__ip_vs_conn_get(cp))
 					goto out;
@@ -366,9 +366,10 @@ struct ip_vs_conn *ip_vs_ct_in_get(const struct ip_vs_conn_param *p)
 		     * p->vaddr is a fwmark */
 		    ip_vs_addr_equal(p->protocol == IPPROTO_IP ? AF_UNSPEC :
 				     p->af, p->vaddr, &cp->vaddr) &&
-		    p->cport == cp->cport && p->vport == cp->vport &&
+		    p->vport == cp->vport && p->cport == cp->cport &&
 		    cp->flags & IP_VS_CONN_F_TEMPLATE &&
-		    p->protocol == cp->protocol) {
+		    p->protocol == cp->protocol &&
+		    ip_vs_conn_net_eq(cp, p->net)) {
 			if (__ip_vs_conn_get(cp))
 				goto out;
 		}
@@ -404,8 +405,8 @@ struct ip_vs_conn *ip_vs_conn_out_get(const struct ip_vs_conn_param *p)
 	rcu_read_lock();
 
 	hlist_for_each_entry_rcu(cp, &ip_vs_conn_tab[hash], c_list) {
-		if (cp->af == p->af &&
-		    p->vport == cp->cport && p->cport == cp->dport &&
+		if (p->vport == cp->cport && p->cport == cp->dport &&
+		    cp->af == p->af &&
 		    ip_vs_addr_equal(p->af, p->vaddr, &cp->caddr) &&
 		    ip_vs_addr_equal(p->af, p->caddr, &cp->daddr) &&
 		    p->protocol == cp->protocol &&
Signature · 1 line
1.7.3.4
author maintainer
#15 lore raw

PATCH v3 15/15 ipvs: avoid kmem_cache_zalloc in ip_vs_conn_new

In reply to Julian Anastasov

We have many fields to set and few to reset,
use kmem_cache_alloc instead to save some cycles.

---
 include/net/ip_vs.h             |   15 +++++++++++++++
 net/netfilter/ipvs/ip_vs_conn.c |   24 +++++++++++++++++++-----
 2 files changed, 34 insertions(+), 5 deletions(-)
include/net/ip_vs.h +15 -0
index 26f104a..5ed8939 100644
@@ -233,6 +233,21 @@ static inline void ip_vs_addr_copy(int af, union nf_inet_addr *dst,
 	dst->ip = src->ip;
 }
 
+static inline void ip_vs_addr_set(int af, union nf_inet_addr *dst,
+				  const union nf_inet_addr *src)
+{
+#ifdef CONFIG_IP_VS_IPV6
+	if (af == AF_INET6) {
+		dst->in6 = src->in6;
+		return;
+	}
+#endif
+	dst->ip = src->ip;
+	dst->all[1] = 0;
+	dst->all[2] = 0;
+	dst->all[3] = 0;
+}
+
 static inline int ip_vs_addr_equal(int af, const union nf_inet_addr *a,
 				   const union nf_inet_addr *b)
 {
net/netfilter/ipvs/ip_vs_conn.c +19 -5
index 416015b..e3e2b4d 100644
@@ -861,7 +861,7 @@ ip_vs_conn_new(const struct ip_vs_conn_param *p,
 	struct ip_vs_proto_data *pd = ip_vs_proto_data_get(p->net,
 							   p->protocol);
 
-	cp = kmem_cache_zalloc(ip_vs_conn_cachep, GFP_ATOMIC);
+	cp = kmem_cache_alloc(ip_vs_conn_cachep, GFP_ATOMIC);
 	if (cp == NULL) {
 		IP_VS_ERR_RL("%s(): no memory\n", __func__);
 		return NULL;
@@ -872,13 +872,13 @@ ip_vs_conn_new(const struct ip_vs_conn_param *p,
 	ip_vs_conn_net_set(cp, p->net);
 	cp->af		   = p->af;
 	cp->protocol	   = p->protocol;
-	ip_vs_addr_copy(p->af, &cp->caddr, p->caddr);
+	ip_vs_addr_set(p->af, &cp->caddr, p->caddr);
 	cp->cport	   = p->cport;
-	ip_vs_addr_copy(p->af, &cp->vaddr, p->vaddr);
+	ip_vs_addr_set(p->af, &cp->vaddr, p->vaddr);
 	cp->vport	   = p->vport;
 	/* proto should only be IPPROTO_IP if d_addr is a fwmark */
-	ip_vs_addr_copy(p->protocol == IPPROTO_IP ? AF_UNSPEC : p->af,
-			&cp->daddr, daddr);
+	ip_vs_addr_set(p->protocol == IPPROTO_IP ? AF_UNSPEC : p->af,
+		       &cp->daddr, daddr);
 	cp->dport          = dport;
 	cp->flags	   = flags;
 	cp->fwmark         = fwmark;
@@ -887,6 +887,10 @@ ip_vs_conn_new(const struct ip_vs_conn_param *p,
 		cp->pe = p->pe;
 		cp->pe_data = p->pe_data;
 		cp->pe_data_len = p->pe_data_len;
+	} else {
+		cp->pe = NULL;
+		cp->pe_data = NULL;
+		cp->pe_data_len = 0;
 	}
 	spin_lock_init(&cp->lock);
 
@@ -897,18 +901,28 @@ ip_vs_conn_new(const struct ip_vs_conn_param *p,
 	 */
 	atomic_set(&cp->refcnt, 1);
 
+	cp->control = NULL;
 	atomic_set(&cp->n_control, 0);
 	atomic_set(&cp->in_pkts, 0);
 
+	cp->packet_xmit = NULL;
+	cp->app = NULL;
+	cp->app_data = NULL;
+	/* reset struct ip_vs_seq */
+	cp->in_seq.delta = 0;
+	cp->out_seq.delta = 0;
+
 	atomic_inc(&ipvs->conn_count);
 	if (flags & IP_VS_CONN_F_NO_CPORT)
 		atomic_inc(&ip_vs_conn_no_cport_cnt);
 
 	/* Bind the connection with a destination server */
+	cp->dest = NULL;
 	ip_vs_bind_dest(cp, dest);
 
 	/* Set its state and timeout */
 	cp->state = 0;
+	cp->old_state = 0;
 	cp->timeout = 3*HZ;
 	cp->sync_endtime = jiffies & ~3UL;
 
Signature · 1 line
1.7.3.4
#16 lore raw
In reply to Julian Anastasov

Hello Julian
Great work!

Quoted 11 lines · Julian Anastasov

This is a first patchset for IPVS optimizations.
Another patchset will address the locking in schedulers and moving the global _bh disabling from LOCAL_OUT to all locks.

All patches are for net-next and Simon can
take them for ipvs-next.

The changes in this patchset eliminate locks
and dst refcnt operations from packet processing by using RCU. There are more details in the patches.

I've gone through all patches and tested them now it looks very promising, all my tests have passed.

Signed-off by: Hans Schillstrom <[email protected]>

Quoted 62 lines

v3:
* in "ipvs: consolidate all dst checks on transmit in one place" preserve original skb dst even for local client, remove the rt_is_input_route and skb->dev check. Call update_pmtu only for local client by providing sk instead of skb.
* in "ipvs: optimize dst usage for real server" use rcu_dereference_protected for __ip_vs_dst_cache_reset instead of rcu_dereference_raw. Use the new skb_dst_set_noref_force func.
* in "ipvs: remove rs_lock by using RCU" prefer the port check in ip_vs_has_real_service
* "ipvs: convert locks used in persistence engines" needs only synchronize_rcu, not rcu_barrier, we do not use rcu callbacks

v2:
* use "net: add skb_dst_set_unref" instead of
"net: add dst_get_noref and refdst_ptr helpers"
* add "ipvs: no need to reroute anymore on DNAT over loopback"
* add "ipvs: do not use skb_share_check"
* add "ipvs: consolidate all dst checks on transmit in one place", so that we can avoid the refdst games in next patch
* after "ipvs: consolidate all dst checks on transmit in one place" "ipvs: optimize dst usage for real server" is simpler and uses the new skb_dst_set_unref function
* extend "ipvs: reorder keys in connection structure" with changes in ip_vs_ct_in_get
* fix "ipvs: avoid kmem_cache_zalloc in ip_vs_conn_new" to use new function ip_vs_addr_set, so that we reset all address fields that are used for hashing by hash_conntrack_raw

Julian Anastasov (15):
net: add skb_dst_set_noref_force
ipvs: avoid routing by TOS for real server
ipvs: prefer NETDEV_DOWN event to free cached dsts
ipvs: convert the IP_VS_XMIT macros to functions
ipvs: rename functions related to dst_cache reset
ipvs: no need to reroute anymore on DNAT over loopback
ipvs: do not use skb_share_check
ipvs: consolidate all dst checks on transmit in one place
ipvs: optimize dst usage for real server
ipvs: convert app locks
ipvs: remove rs_lock by using RCU
ipvs: convert locks used in persistence engines
ipvs: convert connection locking
ipvs: reorder keys in connection structure
ipvs: avoid kmem_cache_zalloc in ip_vs_conn_new

 include/linux/skbuff.h                |   35 ++-
 include/net/ip_vs.h                   |   71 ++-
 net/core/dst.c                        |    9 +-
 net/netfilter/ipvs/ip_vs_app.c        |   27 +-
 net/netfilter/ipvs/ip_vs_conn.c       |  271 +++++----
 net/netfilter/ipvs/ip_vs_core.c       |   16 +-
 net/netfilter/ipvs/ip_vs_ctl.c        |  143 +++---
 net/netfilter/ipvs/ip_vs_ftp.c        |    2 +
 net/netfilter/ipvs/ip_vs_pe.c         |   43 +-
 net/netfilter/ipvs/ip_vs_pe_sip.c     |    1 +
 net/netfilter/ipvs/ip_vs_proto_sctp.c |   18 +-
 net/netfilter/ipvs/ip_vs_proto_tcp.c  |   18 +-
 net/netfilter/ipvs/ip_vs_proto_udp.c  |   19 +-
 net/netfilter/ipvs/ip_vs_xmit.c       | 1046 ++++++++++++++-------------------
 14 files changed, 810 insertions(+), 909 deletions(-)
maintainer
#17 lore raw
Quoted 12 lines · Hans Schillstrom

Hello Julian
Great work!

Quoted 8 lines · Julian Anastasov

This is a first patchset for IPVS optimizations.
Another patchset will address the locking in schedulers and moving the global _bh disabling from LOCAL_OUT to all locks.

All patches are for net-next and Simon can
take them for ipvs-next.

Quoted 8 lines
Quoted 3 lines

The changes in this patchset eliminate locks
and dst refcnt operations from packet processing by using RCU. There are more details in the patches.

I've gone through all patches and tested them now it looks very promising, all my tests have passed.

Signed-off by: Hans Schillstrom <[email protected]>

Thanks, and sorry to be a bit slow.

I will push these to the ipvs-next tree.
I will then post them to netdev and David Miller for review as I think the first patch falls into that realm.

author maintainer
#18 lore raw

Hello,

Quoted 20 lines · Simon Horman
Quoted 19 lines · Hans Schillstrom

Hello Julian
Great work!

Quoted 11 lines · Julian Anastasov

This is a first patchset for IPVS optimizations.
Another patchset will address the locking in schedulers and moving the global _bh disabling from LOCAL_OUT to all locks.

All patches are for net-next and Simon can
take them for ipvs-next.

The changes in this patchset eliminate locks
and dst refcnt operations from packet processing by using RCU. There are more details in the patches.

I've gone through all patches and tested them now it looks very promising, all my tests have passed.

Quoted 7 lines
Quoted 1 line

Signed-off by: Hans Schillstrom <[email protected]>

Thanks, and sorry to be a bit slow.

I will push these to the ipvs-next tree.
I will then post them to netdev and David Miller for review as I think the first patch falls into that realm.

He already acked v2 of this patch and I added his
last comment in v3:

http://marc.info/?l=linux-netdev&m=136294921002680&w=2

I preferred not to disturb netdev with v3, it
has small changes after v2.

Regards

--
Julian Anastasov <[email protected]>

maintainer
#19 lore raw
Quoted 33 lines · Julian Anastasov

Hello,

Quoted 20 lines · Simon Horman
Quoted 19 lines · Hans Schillstrom

Hello Julian
Great work!

Quoted 11 lines · Julian Anastasov

This is a first patchset for IPVS optimizations.
Another patchset will address the locking in schedulers and moving the global _bh disabling from LOCAL_OUT to all locks.

All patches are for net-next and Simon can
take them for ipvs-next.

The changes in this patchset eliminate locks
and dst refcnt operations from packet processing by using RCU. There are more details in the patches.

I've gone through all patches and tested them now it looks very promising, all my tests have passed.

Quoted 7 lines
Quoted 1 line

Signed-off by: Hans Schillstrom <[email protected]>

Thanks, and sorry to be a bit slow.

I will push these to the ipvs-next tree.
I will then post them to netdev and David Miller for review as I think the first patch falls into that realm.

Quoted 7 lines

He already acked v2 of this patch and I added his
last comment in v3:

http://marc.info/?l=linux-netdev&m=136294921002680&w=2

I preferred not to disturb netdev with v3, it
has small changes after v2.

Thanks, and sorry for missing that.