Skip to contentSkip to conversation

inetpeer: Support ipv6 addresses.

PATCH In review David Miller 5 participants 13 messages Nov 29, 2010 – Nov 30, 2010

Subsystems (4)
lore
Share & tools
lore

David Miller

author maintainer
#0 lore raw

PATCH 1/3 inetpeer: Support ipv6 addresses.

---
 include/net/inetpeer.h |   21 ++++++++++++++-
 net/ipv4/inetpeer.c    |   66 ++++++++++++++++++++++++++++++++++++++++-------
 net/ipv4/ip_fragment.c |    2 +-
 net/ipv4/route.c       |    2 +-
 net/ipv4/tcp_ipv4.c    |    6 ++--
 5 files changed, 80 insertions(+), 17 deletions(-)
include/net/inetpeer.h +19 -2
index fe239bf..834f045 100644
@@ -13,10 +13,18 @@
 #include <linux/spinlock.h>
 #include <asm/atomic.h>
 
+typedef struct {
+	union {
+		__be32		a4;
+		__be32		a6[4];
+	};
+	__u16	family;
+} inet_peer_address_t;
+
 struct inet_peer {
 	/* group together avl_left,avl_right,v4daddr to speedup lookups */
 	struct inet_peer __rcu	*avl_left, *avl_right;
-	__be32			v4daddr;	/* peer's address */
+	inet_peer_address_t	daddr;
 	__u32			avl_height;
 	struct list_head	unused;
 	__u32			dtime;		/* the time of last use of not
@@ -42,7 +50,16 @@ struct inet_peer {
 void			inet_initpeers(void) __init;
 
 /* can be called with or without local BH being disabled */
-struct inet_peer	*inet_getpeer(__be32 daddr, int create);
+struct inet_peer	*inet_getpeer(inet_peer_address_t *daddr, int create);
+
+static inline struct inet_peer *inet_getpeer_v4(__be32 v4daddr, int create)
+{
+	inet_peer_address_t daddr;
+
+	daddr.a4 = v4daddr;
+	daddr.family = AF_INET;
+	return inet_getpeer(&daddr, create);
+}
 
 /* can be called from BH context or outside */
 extern void inet_putpeer(struct inet_peer *p);
net/ipv4/inetpeer.c +56 -10

133 lines not on this page · raw message

net/ipv4/ip_fragment.c +1 -1
index 1684408..e6215bd 100644
@@ -141,7 +141,7 @@ static void ip4_frag_init(struct inet_frag_queue *q, void *a)
 	qp->daddr = arg->iph->daddr;
 	qp->user = arg->user;
 	qp->peer = sysctl_ipfrag_max_dist ?
-		inet_getpeer(arg->iph->saddr, 1) : NULL;
+		inet_getpeer_v4(arg->iph->saddr, 1) : NULL;
 }
 
 static __inline__ void ip4_frag_free(struct inet_frag_queue *q)
net/ipv4/route.c +1 -1
index ec2333f..3843c2d 100644
@@ -1289,7 +1289,7 @@ void rt_bind_peer(struct rtable *rt, int create)
 {
 	struct inet_peer *peer;
 
-	peer = inet_getpeer(rt->rt_dst, create);
+	peer = inet_getpeer_v4(rt->rt_dst, create);
 
 	if (peer && cmpxchg(&rt->peer, NULL, peer) != NULL)
 		inet_putpeer(peer);
net/ipv4/tcp_ipv4.c +3 -3
index 69ccbc1..00285fc 100644
@@ -1347,7 +1347,7 @@ int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)
 		    tcp_death_row.sysctl_tw_recycle &&
 		    (dst = inet_csk_route_req(sk, req)) != NULL &&
 		    (peer = rt_get_peer((struct rtable *)dst)) != NULL &&
-		    peer->v4daddr == saddr) {
+		    peer->daddr.a4 == saddr) {
 			inet_peer_refcheck(peer);
 			if ((u32)get_seconds() - peer->tcp_ts_stamp < TCP_PAWS_MSL &&
 			    (s32)(peer->tcp_ts - req->ts_recent) >
@@ -1778,7 +1778,7 @@ int tcp_v4_remember_stamp(struct sock *sk)
 	int release_it = 0;
 
 	if (!rt || rt->rt_dst != inet->inet_daddr) {
-		peer = inet_getpeer(inet->inet_daddr, 1);
+		peer = inet_getpeer_v4(inet->inet_daddr, 1);
 		release_it = 1;
 	} else {
 		if (!rt->peer)
@@ -1804,7 +1804,7 @@ EXPORT_SYMBOL(tcp_v4_remember_stamp);
 
 int tcp_v4_tw_remember_stamp(struct inet_timewait_sock *tw)
 {
-	struct inet_peer *peer = inet_getpeer(tw->tw_daddr, 1);
+	struct inet_peer *peer = inet_getpeer_v4(tw->tw_daddr, 1);
 
 	if (peer) {
 		const struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
Signature · 1 line
1.7.3.2
#1 lore raw
Quoted 1 line · On Mon, 29 Nov 2010 13:44:03 -0800 (PST) David Miller <[email protected]>…

+static inline bool inet_peer_addr_equal(inet_peer_address_t *a, inet_peer_address_t *b)

Let the compiler decide about inline.
Should be const inet_peer_address_t *

author maintainer
#2 lore raw

From: Stephen Hemminger <[email protected]>
Date: Mon, 29 Nov 2010 14:33:57 -0800

Quoted 7 lines
Quoted 1 line · On Mon, 29 Nov 2010 13:44:03 -0800 (PST) David Miller <[email protected]>…

+static inline bool inet_peer_addr_equal(inet_peer_address_t *a, inet_peer_address_t *b)

Let the compiler decide about inline.
Should be const inet_peer_address_t *

Yep, I made those changes thanks.

#3 lore raw
In reply to David Miller

net/ipv4/inetpeer.c static void unlink_from_unused(struct inet_peer *p) · line 168

Quoted 13 lines · David Miller
net/ipv4/inetpeer.c +13 -0
+static inline bool inet_peer_addr_equal(inet_peer_address_t *a, inet_peer_address_t *b)
+{
+	if (a->family == b->family) {
+		switch (a->family) {
+		case AF_INET:
+			if (a->a4 == b->a4)
+				return true;
+			break;
+		case AF_INET6:
+			if (!ipv6_addr_cmp((struct in6_addr *)a,
+					   (struct in6_addr *)b))
+				return true;
+			break;

ipv6_addr_equal() ?

-Brian

author maintainer
#4 lore raw

From: Brian Haley <[email protected]>
Date: Mon, 29 Nov 2010 21:11:55 -0500

Quoted 8 lines · David Miller

net/ipv4/inetpeer.c static void unlink_from_unused(struct inet_peer *p) · line 168

Quoted 5 lines
net/ipv4/inetpeer.c +5 -0
+		case AF_INET6:
+			if (!ipv6_addr_cmp((struct in6_addr *)a,
+					   (struct in6_addr *)b))
+				return true;
+			break;

ipv6_addr_equal() ?

Yep, makes sense to me.

#5 lore raw
In reply to David Miller
Quoted 16 lines · David Miller

---
 include/net/inetpeer.h |   21 ++++++++++++++- net/ipv4/inetpeer.c    |   66 ++++++++++++++++++++++++++++++++++++++++------- net/ipv4/ip_fragment.c |    2 +-
 net/ipv4/route.c       |    2 +-
 net/ipv4/tcp_ipv4.c    |    6 ++--

 5 files changed, 80 insertions(+), 17 deletions(-)
include/net/inetpeer.h +0 -0
index fe239bf..834f045 100644
@@ -13,10 +13,18 @@
  #include <linux/spinlock.h>

include/net/inetpeer.h line 23

Quoted 10 lines
include/net/inetpeer.h +8 -0
  #include <asm/atomic.h>
 
+typedef struct {
+       union {
+               __be32          a4;
+               __be32          a6[4];
+       };
+       __u16   family;
+} inet_peer_address_t;
+

This makes inet_peer large even though ipv6 isn't enabled.

include/net/inetpeer.h line 27

Quoted 5 lines
include/net/inetpeer.h +1 -1
  struct inet_peer {
        /* group together avl_left,avl_right,v4daddr to speedup lookups */
        struct inet_peer __rcu  *avl_left, *avl_right;
-       __be32                  v4daddr;        /* peer's address */
+       inet_peer_address_t     daddr;

I have thought about converting this AVL tree to rbtree. When I saw the comment above, I gave it up, because rbtree makes this structure bigger. If ipv6 support is added, I think it is time to turn to rbtree. :)

Quoted 232 lines
include/net/inetpeer.h +0 -0
        __u32                   avl_height;
        struct list_head        unused;

       __u32                   dtime;          /* the time of last use of not

include/net/inetpeer.h +10 -1
@@ -42,7 +50,16 @@ struct inet_peer {
  void                   inet_initpeers(void) __init;
 
  /* can be called with or without local BH being disabled */
-struct inet_peer       *inet_getpeer(__be32 daddr, int create);
+struct inet_peer       *inet_getpeer(inet_peer_address_t *daddr, int create);
+
+static inline struct inet_peer *inet_getpeer_v4(__be32 v4daddr, int create)
+{
+       inet_peer_address_t daddr;
+
+       daddr.a4 = v4daddr;
+       daddr.family = AF_INET;
+       return inet_getpeer(&daddr, create);
+}
 
  /* can be called from BH context or outside */
  extern void inet_putpeer(struct inet_peer *p);
net/ipv4/inetpeer.c +49 -5
index 9e94d7c..39f18c5 100644
@@ -19,6 +19,7 @@
  #include <linux/net.h>
  #include <net/ip.h>
  #include <net/inetpeer.h>
+#include <net/ipv6.h>
 
  /*
  *  Theory of operations.
@@ -152,11 +153,53 @@ static void unlink_from_unused(struct inet_peer *p)
        }
  }
 
+static inline bool inet_peer_addr_equal(inet_peer_address_t *a, inet_peer_address_t *b)
+{
+       if (a->family == b->family) {
+               switch (a->family) {
+               case AF_INET:
+                       if (a->a4 == b->a4)
+                               return true;
+                       break;
+               case AF_INET6:
+                       if (!ipv6_addr_cmp((struct in6_addr *)a,
+                                          (struct in6_addr *)b))
+                               return true;
+                       break;
+               default:
+                       break;
+               }
+       }
+       return false;
+}
+
+static inline u32 inet_peer_key(inet_peer_address_t *a)
+{
+       u32 key;
+
+       switch (a->family) {
+       case AF_INET:
+               key = (__force __u32) a->a4;
+               break;
+       case AF_INET6:
+               key = ((__force __u32)a->a6[0] ^
+                      (__force __u32)a->a6[1] ^
+                      (__force __u32)a->a6[2] ^
+                      (__force __u32)a->a6[3]);
+               break;
+       default:
+               key = 0;
+               break;
+       }
+       return key;
+}
+
  /*
  * Called with local BH disabled and the pool lock held.
  */
  #define lookup(_daddr, _stack)                                         \
  ({                                                             \
+       u32 key = inet_peer_key(_daddr);                        \
        struct inet_peer *u;                                    \
        struct inet_peer __rcu **v;                             \
                                                                \
@@ -165,9 +208,9 @@ static void unlink_from_unused(struct inet_peer *p)
        for (u = rcu_dereference_protected(peers.root,          \
                        lockdep_is_held(&peers.lock));          \
             u != peer_avl_empty; ) {                           \
-               if (_daddr == u->v4daddr)                       \
+               if (inet_peer_addr_equal(_daddr, &u->daddr))    \
                        break;                                  \
-               if ((__force __u32)_daddr < (__force __u32)u->v4daddr)  \
+               if (key < inet_peer_key(&u->daddr))             \
                        v = &u->avl_left;                       \
                else                                            \
                        v = &u->avl_right;                      \
@@ -185,13 +228,14 @@ static void unlink_from_unused(struct inet_peer *p)
  * But every pointer we follow is guaranteed to be valid thanks to RCU.
  * We exit from this function if number of links exceeds PEER_MAXDEPTH
  */
-static struct inet_peer *lookup_rcu_bh(__be32 daddr)
+static struct inet_peer *lookup_rcu_bh(inet_peer_address_t *daddr)
  {
        struct inet_peer *u = rcu_dereference_bh(peers.root);
+       u32 key = inet_peer_key(daddr);
        int count = 0;
 
        while (u != peer_avl_empty) {
-               if (daddr == u->v4daddr) {
+               if (inet_peer_addr_equal(daddr, &u->daddr)) {
                        /* Before taking a reference, check if this entry was
                         * deleted, unlink_from_pool() sets refcnt=-1 to make
                         * distinction between an unused entry (refcnt=0) and
@@ -201,7 +245,7 @@ static struct inet_peer *lookup_rcu_bh(__be32 daddr)
                                u = NULL;
                        return u;
                }
-               if ((__force __u32)daddr < (__force __u32)u->v4daddr)
+               if (key < inet_peer_key(&u->daddr))
                        u = rcu_dereference_bh(u->avl_left);

               else
                       u = rcu_dereference_bh(u->avl_right);

net/ipv4/inetpeer.c +1 -1
@@ -353,7 +397,7 @@ static void unlink_from_pool(struct inet_peer *p)
        if (atomic_cmpxchg(&p->refcnt, 1, -1) == 1) {
                struct inet_peer __rcu **stack[PEER_MAXDEPTH];
                struct inet_peer __rcu ***stackptr, ***delp;
-               if (lookup(p->v4daddr, stack) != p)
+               if (lookup(&p->daddr, stack) != p)
                        BUG();

               delp = stackptr - 1; /* *delp[0] == p */ if (p->avl_left == peer_avl_empty_rcu) {

net/ipv4/inetpeer.c +6 -4
@@ -366,7 +410,7 @@ static void unlink_from_pool(struct inet_peer *p)
                        BUG_ON(rcu_dereference_protected(*stackptr[-1],
                                        lockdep_is_held(&peers.lock)) != t);
                        **--stackptr = t->avl_left;
-                       /* t is removed, t->v4daddr > x->v4daddr for any
+                       /* t is removed, t->daddr > x->daddr for any
                         * x in p->avl_left subtree.
                         * Put t in the old place of p. */
                        RCU_INIT_POINTER(*delp[0], t);
@@ -433,7 +477,7 @@ static int cleanup_once(unsigned long ttl)
  }
 
  /* Called with or without local BH being disabled. */
-struct inet_peer *inet_getpeer(__be32 daddr, int create)
+struct inet_peer *inet_getpeer(inet_peer_address_t *daddr, int create)
  {
        struct inet_peer *p;
        struct inet_peer __rcu **stack[PEER_MAXDEPTH], ***stackptr;
@@ -467,10 +511,12 @@ struct inet_peer *inet_getpeer(__be32 daddr, int create)
        }
        p = create ? kmem_cache_alloc(peer_cachep, GFP_ATOMIC) : NULL;
        if (p) {
-               p->v4daddr = daddr;
+               p->daddr = *daddr;
                atomic_set(&p->refcnt, 1);
                atomic_set(&p->rid, 0);
-               atomic_set(&p->ip_id_count, secure_ip_id(daddr));
+               atomic_set(&p->ip_id_count,
+                          (daddr->family == AF_INET) ?
+                          secure_ip_id(daddr->a4) : 0);
                p->tcp_ts_stamp = 0;
                INIT_LIST_HEAD(&p->unused);
 
net/ipv4/ip_fragment.c +0 -0
index 1684408..e6215bd 100644
@@ -141,7 +141,7 @@ static void ip4_frag_init(struct inet_frag_queue *q, void *a)
        qp->daddr = arg->iph->daddr;
        qp->user = arg->user;

       qp->peer = sysctl_ipfrag_max_dist ?
-               inet_getpeer(arg->iph->saddr, 1) : NULL;
+               inet_getpeer_v4(arg->iph->saddr, 1) : NULL; }

 static __inline__ void ip4_frag_free(struct inet_frag_queue *q)

net/ipv4/route.c +1 -1
index ec2333f..3843c2d 100644
@@ -1289,7 +1289,7 @@ void rt_bind_peer(struct rtable *rt, int create)
  {
        struct inet_peer *peer;
 
-       peer = inet_getpeer(rt->rt_dst, create);
+       peer = inet_getpeer_v4(rt->rt_dst, create);
 
        if (peer && cmpxchg(&rt->peer, NULL, peer) != NULL)
                inet_putpeer(peer);
net/ipv4/tcp_ipv4.c +0 -0
index 69ccbc1..00285fc 100644
@@ -1347,7 +1347,7 @@ int tcp_v4_conn_request(struct sock *sk, struct sk_buff *skb)

                   tcp_death_row.sysctl_tw_recycle &&
                   (dst = inet_csk_route_req(sk, req)) != NULL && (peer = rt_get_peer((struct rtable *)dst)) != NULL &&
-                   peer->v4daddr == saddr) {
+                   peer->daddr.a4 == saddr) { inet_peer_refcheck(peer);
                       if ((u32)get_seconds() - peer->tcp_ts_stamp < TCP_PAWS_MSL && (s32)(peer->tcp_ts - req->ts_recent) >

net/ipv4/tcp_ipv4.c +2 -3
@@ -1778,7 +1778,7 @@ int tcp_v4_remember_stamp(struct sock *sk)
        int release_it = 0;
 
        if (!rt || rt->rt_dst != inet->inet_daddr) {
-               peer = inet_getpeer(inet->inet_daddr, 1);
+               peer = inet_getpeer_v4(inet->inet_daddr, 1);
                release_it = 1;
        } else {
                if (!rt->peer)
@@ -1804,7 +1804,7 @@ EXPORT_SYMBOL(tcp_v4_remember_stamp);
 
  int tcp_v4_tw_remember_stamp(struct inet_timewait_sock *tw)
  {
-       struct inet_peer *peer = inet_getpeer(tw->tw_daddr, 1);
+       struct inet_peer *peer = inet_getpeer_v4(tw->tw_daddr, 1);
 
        if (peer) {
                const struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
--

1.7.3.2

--
To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Signature · 2 lines
Regards,
Changli Gao([email protected])
author maintainer
#6 lore raw

From: Changli Gao <[email protected]>
Date: Tue, 30 Nov 2010 10:33:49 +0800

Quoted 4 lines

I have thought about converting this AVL tree to rbtree. When I saw the comment above, I gave it up, because rbtree makes this structure bigger. If ipv6 support is added, I think it is time to turn to rbtree. :)

If it takes size over 128 bytes, it is probably still a bad idea. Right now it is just under 128.

#7 lore raw
Quoted 11 lines · David Miller

From: Changli Gao <[email protected]>
Date: Tue, 30 Nov 2010 10:33:49 +0800

Quoted 4 lines

I have thought about converting this AVL tree to rbtree. When I saw the comment above, I gave it up, because rbtree makes this structure bigger. If ipv6 support is added, I think it is time to turn to rbtree. :)

If it takes size over 128 bytes, it is probably still a bad idea. Right now it is just under 128.

Why 128? The current size of inet_peer is 64 bytes, and after your ipv6 patch, it is just 80. And, inet_peer is allocated from its own mem_cache, so the size of the memory for it should not be aligned to 2^n.

If rbtree is used instead, extra 8 bytes will be needed, and the size of inet_peer will be 88.

Signature · 2 lines
Regards,
Changli Gao([email protected])
author maintainer
#8 lore raw

From: Changli Gao <[email protected]>
Date: Tue, 30 Nov 2010 12:51:08 +0800

Quoted 9 lines · David Miller
Quoted 8 lines

From: Changli Gao <[email protected]>
Date: Tue, 30 Nov 2010 10:33:49 +0800

Quoted 4 lines

I have thought about converting this AVL tree to rbtree. When I saw the comment above, I gave it up, because rbtree makes this structure bigger. If ipv6 support is added, I think it is time to turn to rbtree. :)

Quoted 8 lines
Quoted 3 lines

If it takes size over 128 bytes, it is probably still a bad idea. Right now it is just under 128.

Why 128? The current size of inet_peer is 64 bytes, and after your ipv6 patch, it is just 80. And, inet_peer is allocated from its own mem_cache, so the size of the memory for it should not be aligned to 2^n.

Sorry, I was taking into the consideration other work I am doing which will move all of the routing metrics into inet_peer as well.

With ipv6 address support, it fits perfectly into 128 bytes on 64-bit.

#9 lore raw
Quoted 21 lines · David Miller

From: Changli Gao <[email protected]>
Date: Tue, 30 Nov 2010 12:51:08 +0800

Quoted 9 lines · David Miller
Quoted 8 lines

From: Changli Gao <[email protected]>
Date: Tue, 30 Nov 2010 10:33:49 +0800

Quoted 4 lines

I have thought about converting this AVL tree to rbtree. When I saw the comment above, I gave it up, because rbtree makes this structure bigger. If ipv6 support is added, I think it is time to turn to rbtree. :)

Quoted 8 lines
Quoted 3 lines

If it takes size over 128 bytes, it is probably still a bad idea. Right now it is just under 128.

Why 128? The current size of inet_peer is 64 bytes, and after your ipv6 patch, it is just 80. And, inet_peer is allocated from its own mem_cache, so the size of the memory for it should not be aligned to 2^n.

Quoted 6 lines

Sorry, I was taking into the consideration other work I am doing which will move all of the routing metrics into inet_peer as well.

With ipv6 address support, it fits perfectly into 128 bytes on 64-bit.
--

Its a bit early in the morning here, I must confess I dont yet understand your patch David :)

As we use a tree, why not using two different trees for ipv4 / ipv6 ?

I dont understand how computing a 32bit key (sort of hash key) is going to help when hash collision happens, with an avl tree.

For Changli : AVL tree was OK for RCU conversion, I am not sure about RBtree yet.

Either version of tree (AVL/rbtree) will be expensive to use if depth is big (With 2 millions entries, depth is going to be very big). I understand you want to get rid of route cache ?

author maintainer
#10 lore raw

From: Eric Dumazet <[email protected]>
Date: Tue, 30 Nov 2010 06:42:16 +0100

Quoted 4 lines

Its a bit early in the morning here, I must confess I dont yet understand your patch David :)

As we use a tree, why not using two different trees for ipv4 / ipv6 ?

The "key" just creates a natural ordering in the tree, it's almost arbitrary except that it must distribute well amongst the entries.

I currently don't see any reason to make two trees right now.

Quoted 5 lines

I dont understand how computing a 32bit key (sort of hash key) is going to help when hash collision happens, with an avl tree. Either version of tree (AVL/rbtree) will be expensive to use if depth is big (With 2 millions entries, depth is going to be very big). I understand you want to get rid of route cache ?

Do we plan to talk to 2 million unique destinations and have active non-default metrics for each one of them very often?

inet_peer entries will only get created when we need to make non-default metric settings for a specific destination address.

See that's the thing, it's scope is so much smaller than the existing routing cache. It's only going to be used in limited if not controlled cases.

#11 lore raw
Quoted 11 lines · David Miller

From: Eric Dumazet <[email protected]>
Date: Tue, 30 Nov 2010 06:42:16 +0100

Quoted 4 lines

Its a bit early in the morning here, I must confess I dont yet understand your patch David :)

As we use a tree, why not using two different trees for ipv4 / ipv6 ?

The "key" just creates a natural ordering in the tree, it's almost arbitrary except that it must distribute well amongst the entries.

Hmm. AVL search must take a decision, take the left or the right path.

if current key is equal, which path do you take ?

@@ -165,9 +208,9 @@ static void unlink_from_unused(struct inet_peer *p)

Review

Quoted 11 lines
(quoted diff) +2 -2
 	for (u = rcu_dereference_protected(peers.root,		\
 			lockdep_is_held(&peers.lock));		\
 	     u != peer_avl_empty; ) {				\
-		if (_daddr == u->v4daddr)			\
+		if (inet_peer_addr_equal(_daddr, &u->daddr))	\
 			break;					\
-		if ((__force __u32)_daddr < (__force __u32)u->v4daddr)	\
+		if (key < inet_peer_key(&u->daddr))		\
 			v = &u->avl_left;			\
 		else						\
 			v = &u->avl_right;			\

Apparently you take the right one, you may miss the target if its on the left path ?

Quoted 3 lines

I currently don't see any reason to make two trees right now.

Cost of a tree is one pointer, and ipv4 search would be faster if we use different search functions.

Quoted 9 lines
Quoted 5 lines

I dont understand how computing a 32bit key (sort of hash key) is going to help when hash collision happens, with an avl tree. Either version of tree (AVL/rbtree) will be expensive to use if depth is big (With 2 millions entries, depth is going to be very big). I understand you want to get rid of route cache ?

Do we plan to talk to 2 million unique destinations and have active non-default metrics for each one of them very often?

Quoted 7 lines

inet_peer entries will only get created when we need to make non-default metric settings for a specific destination address.

See that's the thing, it's scope is so much smaller than the existing routing cache. It's only going to be used in limited if not controlled cases.

OK good :)

I have no idea how many addresses have non default metric settings.

Do you know how to make an estimation on a server ?

author maintainer
#12 lore raw

From: Eric Dumazet <[email protected]>
Date: Tue, 30 Nov 2010 07:11:19 +0100

Quoted 9 lines · David Miller
Quoted 8 lines

From: Eric Dumazet <[email protected]>
Date: Tue, 30 Nov 2010 06:42:16 +0100

Quoted 4 lines

Its a bit early in the morning here, I must confess I dont yet understand your patch David :)

As we use a tree, why not using two different trees for ipv4 / ipv6 ?

Quoted 7 lines
Quoted 3 lines

The "key" just creates a natural ordering in the tree, it's almost arbitrary except that it must distribute well amongst the entries.

Hmm. AVL search must take a decision, take the left or the right path.

if current key is equal, which path do you take ?

Right. :-) But yes there is some error that needs to be handled in that equal keys can lead to imporper paths as you showed.

I guess I do need to make a seperate tree, how disappointing.

Anyways I will work on fixing this.

Quoted 1 line

Do you know how to make an estimation on a server ?

No way to do it straightforward I guess because right now TCP metrics key on all sorts of rediculious things like source address, TOS, and other crap as a consequence of how the routing cache works.

So much replicated information and wasted memory.