Skip to contentSkip to conversation

inet_hash_connect: source port allocation

DISCUSSION John Haxby 3 participants 7 messages Nov 29, 2010

Subsystems (3)
lore
Share & tools
lore

John Haxby

author
#0 lore raw

Hello,

Please forgive me if this is a stupid question, but is there any particular reason why the source port allocation in __inet_hash_connect() shouldn't use the same random allocation that inet_csk_get_port() uses? The latter, of course, is used when bind() doesn't specify a source port but the implicit "bind" for a connect() gets its port allocated by __inet_hash_connect().

jch

#1 lore raw
Quoted 10 lines · John Haxby

Hello,

Please forgive me if this is a stupid question, but is there any particular reason why the source port allocation in __inet_hash_connect() shouldn't use the same random allocation that inet_csk_get_port() uses? The latter, of course, is used when bind() doesn't specify a source port but the implicit "bind" for a connect() gets its port allocated by __inet_hash_connect().

jch

autobind vs bind

bind() gives more information, like local address (if any)

autobind(), we dont know local address, it'll be chose later by routing.

author
#2 lore raw
Quoted 10 lines · Eric Dumazet
Quoted 9 lines · John Haxby

Hello,

Please forgive me if this is a stupid question, but is there any particular reason why the source port allocation in __inet_hash_connect() shouldn't use the same random allocation that inet_csk_get_port() uses? The latter, of course, is used when bind() doesn't specify a source port but the implicit "bind" for a connect() gets its port allocated by __inet_hash_connect().

Quoted 6 lines
Quoted 1 line

jch

autobind vs bind

bind() gives more information, like local address (if any)

autobind(), we dont know local address, it'll be chose later by routing.

Sorry, I think I phrased my question badly.

inet_csk_get_port() starts its search for a free port with

    smallest_rover = rover = net_random() % remaining + low;

whereas __inet_hash_connect() basically misses out that call to net_random() so you get a predictable port number.

Is there any good reason why that is the case?

jch

#3 lore raw
Quoted 11 lines · John Haxby

Sorry, I think I phrased my question badly.

inet_csk_get_port() starts its search for a free port with

     smallest_rover = rover = net_random() % remaining + low;

whereas __inet_hash_connect() basically misses out that call to net_random() so you get a predictable port number.

Is there any good reason why that is the case?

It seems random select was done at bind() time only in commit 6df716340da3a6f ([TCP/DCCP]: Randomize port selection)

It probably should be done in autobind too.

#4 lore raw
Quoted 11 lines · Eric Dumazet
Quoted 9 lines · John Haxby

Sorry, I think I phrased my question badly.

inet_csk_get_port() starts its search for a free port with

     smallest_rover = rover = net_random() % remaining + low;

whereas __inet_hash_connect() basically misses out that call to net_random() so you get a predictable port number.

Quoted 9 lines
Quoted 2 lines

Is there any good reason why that is the case?

It seems random select was done at bind() time only in commit 6df716340da3a6f ([TCP/DCCP]: Randomize port selection)

It probably should be done in autobind too.

I'll test following patch :

net/ipv4/inet_hashtables.c +9 -10
index 1b344f3..65c3702 100644
@@ -466,20 +466,18 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 	int twrefcnt = 1;
 
 	if (!snum) {
-		int i, remaining, low, high, port;
-		static u32 hint;
-		u32 offset = hint + port_offset;
+		int remaining, low, high, port;
 		struct hlist_node *node;
 		struct inet_timewait_sock *tw = NULL;
 
 		inet_get_local_port_range(&low, &high);
 		remaining = (high - low) + 1;
+		port = net_random() % remaining + low;
 
 		local_bh_disable();
-		for (i = 1; i <= remaining; i++) {
-			port = low + (i + offset) % remaining;
+		do {
 			if (inet_is_reserved_local_port(port))
-				continue;
+				goto next_nolock;
 			head = &hinfo->bhash[inet_bhashfn(net, port,
 					hinfo->bhash_size)];
 			spin_lock(&head->lock);
@@ -510,16 +508,17 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 			tb->fastreuse = -1;
 			goto ok;
 
-		next_port:
+next_port:
 			spin_unlock(&head->lock);
-		}
+next_nolock:
+			if (++port > high)
+				port = low;
+		} while (--remaining > 0);
 		local_bh_enable();
 
 		return -EADDRNOTAVAIL;
 
 ok:
-		hint += i;
-
 		/* Head lock still held and bh's disabled */
 		inet_bind_hash(sk, tb, port);
 		if (sk_unhashed(sk)) {
#5 lore raw
Quoted 15 lines · Eric Dumazet
Quoted 14 lines · Eric Dumazet
Quoted 11 lines · John Haxby

Sorry, I think I phrased my question badly.

inet_csk_get_port() starts its search for a free port with

     smallest_rover = rover = net_random() % remaining + low;

whereas __inet_hash_connect() basically misses out that call to net_random() so you get a predictable port number.

Is there any good reason why that is the case?

Quoted 8 lines
Quoted 6 lines

It seems random select was done at bind() time only in commit 6df716340da3a6f ([TCP/DCCP]: Randomize port selection)

It probably should be done in autobind too.

I'll test following patch :

Oh well, forget this, there is something about inet_sk_port_offset() using secure_ipv4_port_ephemeral()

We want to avoid reusing same port too fast.

http://www.tcpipguide.com/free/t_TCPIPClientEphemeralPortsandClientServerApplicatio-2.htm

Port is predictable only for same destination, and if no other connections are attempted by other threads.

#6 lore raw
In reply to Eric Dumazet
Quoted 66 lines · On Mon, 29 Nov 2010 20:07:35 +0100 Eric Dumazet <[email protected]> wrote:
Quoted 11 lines · Eric Dumazet
Quoted 9 lines · John Haxby

Sorry, I think I phrased my question badly.

inet_csk_get_port() starts its search for a free port with

     smallest_rover = rover = net_random() % remaining + low;

whereas __inet_hash_connect() basically misses out that call to net_random() so you get a predictable port number.

Quoted 9 lines
Quoted 2 lines

Is there any good reason why that is the case?

It seems random select was done at bind() time only in commit 6df716340da3a6f ([TCP/DCCP]: Randomize port selection)

It probably should be done in autobind too.

I'll test following patch :

net/ipv4/inet_hashtables.c +9 -8
index 1b344f3..65c3702 100644
@@ -466,20 +466,18 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 	int twrefcnt = 1;
 
 	if (!snum) {
-		int i, remaining, low, high, port;
-		static u32 hint;
-		u32 offset = hint + port_offset;
+		int remaining, low, high, port;
 		struct hlist_node *node;
 		struct inet_timewait_sock *tw = NULL;
 
 		inet_get_local_port_range(&low, &high);
 		remaining = (high - low) + 1;
+		port = net_random() % remaining + low;
 
 		local_bh_disable();
-		for (i = 1; i <= remaining; i++) {
-			port = low + (i + offset) % remaining;
+		do {
 			if (inet_is_reserved_local_port(port))
-				continue;
+				goto next_nolock;
 			head = &hinfo->bhash[inet_bhashfn(net, port,
 					hinfo->bhash_size)];
 			spin_lock(&head->lock);
@@ -510,16 +508,17 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
 			tb->fastreuse = -1;
 			goto ok;
 
-		next_port:
+next_port:
 			spin_unlock(&head->lock);
-		}
+next_nolock:
+			if (++port > high)
+				port = low;
+		} while (--remaining > 0);

net/ipv4/inet_hashtables.c int __inet_hash_connect(struct inet_timewait_death_row *death_row, · line 524

Quoted 10 lines
net/ipv4/inet_hashtables.c +0 -2
 		local_bh_enable();
 
 		return -EADDRNOTAVAIL;
 
 ok:
-		hint += i;
-
 		/* Head lock still held and bh's disabled */
 		inet_bind_hash(sk, tb, port);
 		if (sk_unhashed(sk)) {

The original algorithm works better than uses if the port space is small and being reused rapidly. Because the hint in the old algorithm is sequential ports get used up sequentially.

You should look a the port randomization RFC. The earlier versions of the RFC were better before the BSD guys started putting in their non-scalable algorithms :-)