[PATCH] tcp: Fix inet_bind2_bucket_match_addr_any() regression

[PATCH] tcp: Fix inet_bind2_bucket_match_addr_any() regression

[Jianguo Wu]

From: Jianguo Wu <wujianguo@chinatelecom.cn>

If we bind() a TCPv4 socket to 0.0.0.0:8090, then bind() a TCPv6(ipv6only) socket
to :::8090, both without SO_REUSEPORT, then bind() 127.0.0.1:8090, it should fail
but now succeeds. like this:
  tcp        0      0 127.0.0.1:8090          0.0.0.0:*               LISTEN
  tcp        0      0 0.0.0.0:8090            0.0.0.0:*               LISTEN
  tcp6       0      0 :::8090                 :::*                    LISTEN

bind() 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.

But if we bind() a TCPv6(ipv6only) socket to :::8090 first, then  bind() a TCPv4
socket to 0.0.0.0:8090, then bind() 127.0.0.1:8090, 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.

When bind() 127.0.0.1:8090, inet_bind2_bucket_match_addr_any() will return true as tb->addr_type == IPV6_ADDR_ANY,
and tb is refer to the TCPv6 socket(:::8090), then inet_bhash2_conflict() return false, That is, there is no conflict,
so bind() succeeds.

  inet_bhash2_addr_any_conflict()
  {
	inet_bind_bucket_for_each(tb2, &head2->chain)
		// tb2 is IPv6
		if (inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
			break;

	// inet_bhash2_conflict() return false
	if (tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
                                reuseport_ok)) {
		spin_unlock(&head2->lock);
		return true;
	}

  }

Fixes: 5a22bba13d01 ("tcp: Save address type in inet_bind2_bucket.")
---
 net/ipv4/inet_hashtables.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index 7498af320164..3eeaca8a113f 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -830,8 +830,8 @@ bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const
 		return false;

 #if IS_ENABLED(CONFIG_IPV6)
-	if (tb->addr_type == IPV6_ADDR_ANY)
-		return true;
+	if (sk->sk_family == AF_INET6)
+		return tb->addr_type == IPV6_ADDR_ANY;

 	if (tb->addr_type != IPV6_ADDR_MAPPED)
 		return false;
-- 
1.8.3.1

Re: [PATCH] tcp: Fix inet_bind2_bucket_match_addr_any() regression

[Kuniyuki Iwashima]

Hi,

Thanks for the patch.

From: Jianguo Wu <wujianguo106@163.com>
Date: Thu, 21 Mar 2024 11:02:36 +0800
> From: Jianguo Wu <wujianguo@chinatelecom.cn>
> 
> If we bind() a TCPv4 socket to 0.0.0.0:8090, then bind() a TCPv6(ipv6only) socket

Please wrap each line at <75 characters except for logs/output like below.


> to :::8090, both without SO_REUSEPORT, then bind() 127.0.0.1:8090, it should fail

[::]:8090 is easier to read and the recommended way.
https://datatracker.ietf.org/doc/html/rfc5952#section-6

But please keep the netstat output as is.

> but now succeeds. like this:
>   tcp        0      0 127.0.0.1:8090          0.0.0.0:*               LISTEN
>   tcp        0      0 0.0.0.0:8090            0.0.0.0:*               LISTEN
>   tcp6       0      0 :::8090                 :::*                    LISTEN
> 
> bind() 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.

What do you mean by all fail ?
At least, [::1]:8090 would fail with the current code in this case.


> But if we bind() a TCPv6(ipv6only) socket to :::8090 first, then  bind() a TCPv4
> socket to 0.0.0.0:8090, then bind() 127.0.0.1:8090, 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.
> 
> When bind() 127.0.0.1:8090, inet_bind2_bucket_match_addr_any() will return true as tb->addr_type == IPV6_ADDR_ANY,

Let's use tb2 here for inet_bind2_bucket.. yes it's not consistent
in some functions like inet_bind2_bucket_match_addr_any() though.


> and tb is refer to the TCPv6 socket(:::8090), then inet_bhash2_conflict() return false, That is, there is no conflict,

Also make it clear that the TCPv6 socket is ipv6only one.


> so bind() succeeds.
> 
>   inet_bhash2_addr_any_conflict()
>   {
> 	inet_bind_bucket_for_each(tb2, &head2->chain)
> 		// tb2 is IPv6
> 		if (inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
> 			break;
> 
> 	// inet_bhash2_conflict() return false
> 	if (tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
>                                 reuseport_ok)) {
> 		spin_unlock(&head2->lock);
> 		return true;
> 	}
> 
>   }
> 
> Fixes: 5a22bba13d01 ("tcp: Save address type in inet_bind2_bucket.")

This is not the commit that introduced the regression.

Also, you need Signed-off-by tag here.


> ---
>  net/ipv4/inet_hashtables.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
> index 7498af320164..3eeaca8a113f 100644
> --- a/net/ipv4/inet_hashtables.c
> +++ b/net/ipv4/inet_hashtables.c
> @@ -830,8 +830,8 @@ bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const
>  		return false;
> 
>  #if IS_ENABLED(CONFIG_IPV6)
> -	if (tb->addr_type == IPV6_ADDR_ANY)
> -		return true;
> +	if (sk->sk_family == AF_INET6)
> +		return tb->addr_type == IPV6_ADDR_ANY;

This fix is not correct and will break v4-mapped-v6 address cases.
You can run bind_wildcard under the selftest directory.

Probably we need v6_only bit in tb2 and should add some test cases
in the selftest.


> 
>  	if (tb->addr_type != IPV6_ADDR_MAPPED)
>  		return false;

Re: [PATCH] tcp: Fix inet_bind2_bucket_match_addr_any() regression

[Jianguo Wu]

Hi Kuniyuki,
Thanks for your reply!

On 2024/3/21 12:55, Kuniyuki Iwashima wrote:
> Hi,
> 
> Thanks for the patch.
> 
> From: Jianguo Wu <wujianguo106@163.com>
> Date: Thu, 21 Mar 2024 11:02:36 +0800
>> From: Jianguo Wu <wujianguo@chinatelecom.cn>
>>
>> If we bind() a TCPv4 socket to 0.0.0.0:8090, then bind() a TCPv6(ipv6only) socket
> 
> Please wrap each line at <75 characters except for logs/output like below.
> 
OK.
> 
>> to :::8090, both without SO_REUSEPORT, then bind() 127.0.0.1:8090, it should fail
> 
> [::]:8090 is easier to read and the recommended way.
> https://datatracker.ietf.org/doc/html/rfc5952#section-6
> 
> But please keep the netstat output as is.
> 
>> but now succeeds. like this:
>>   tcp        0      0 127.0.0.1:8090          0.0.0.0:*               LISTEN
>>   tcp        0      0 0.0.0.0:8090            0.0.0.0:*               LISTEN
>>   tcp6       0      0 :::8090                 :::*                    LISTEN
>>
>> bind() 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.
> 
> What do you mean by all fail ?
> At least, [::1]:8090 would fail with the current code in this case.
In my test, 127.0.0.1:8090  succeeds, 0.0.0.0:8090, [::]:8090 and [::1]:8090 are all fail

> 
> 
>> But if we bind() a TCPv6(ipv6only) socket to :::8090 first, then  bind() a TCPv4
>> socket to 0.0.0.0:8090, then bind() 127.0.0.1:8090, 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.
>>
>> When bind() 127.0.0.1:8090, inet_bind2_bucket_match_addr_any() will return true as tb->addr_type == IPV6_ADDR_ANY,
> 
> Let's use tb2 here for inet_bind2_bucket.. yes it's not consistent
> in some functions like inet_bind2_bucket_match_addr_any() though.
yes, inet_bind2_bucket_match_addr_any() use tb, so I use tb here.

> 
> 
>> and tb is refer to the TCPv6 socket(:::8090), then inet_bhash2_conflict() return false, That is, there is no conflict,
> 
> Also make it clear that the TCPv6 socket is ipv6only one.
> 
> 
>> so bind() succeeds.
>>
>>   inet_bhash2_addr_any_conflict()
>>   {
>> 	inet_bind_bucket_for_each(tb2, &head2->chain)
>> 		// tb2 is IPv6
>> 		if (inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
>> 			break;
>>
>> 	// inet_bhash2_conflict() return false
>> 	if (tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
>>                                 reuseport_ok)) {
>> 		spin_unlock(&head2->lock);
>> 		return true;
>> 	}
>>
>>   }
>>
>> Fixes: 5a22bba13d01 ("tcp: Save address type in inet_bind2_bucket.")
> 
> This is not the commit that introduced the regression.
I will remove this.
> 
> Also, you need Signed-off-by tag here.
OK.
> 
> 
>> ---
>>  net/ipv4/inet_hashtables.c | 4 ++--
>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>
>> diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
>> index 7498af320164..3eeaca8a113f 100644
>> --- a/net/ipv4/inet_hashtables.c
>> +++ b/net/ipv4/inet_hashtables.c
>> @@ -830,8 +830,8 @@ bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const
>>  		return false;
>>
>>  #if IS_ENABLED(CONFIG_IPV6)
>> -	if (tb->addr_type == IPV6_ADDR_ANY)
>> -		return true;
>> +	if (sk->sk_family == AF_INET6)
>> +		return tb->addr_type == IPV6_ADDR_ANY;
> 
> This fix is not correct and will break v4-mapped-v6 address cases.
> You can run bind_wildcard under the selftest directory.
>> Probably we need v6_only bit in tb2 and should add some test cases
> in the selftest.
How about this?
I add a new field ipv6_only to struct inet_bind2_bucket{}

diff --git a/include/net/inet_hashtables.h b/include/net/inet_hashtables.h
index 7f1b38458743..fb7c250a663b 100644
--- a/include/net/inet_hashtables.h
+++ b/include/net/inet_hashtables.h
@@ -96,6 +96,7 @@ struct inet_bind2_bucket {
 	int			l3mdev;
 	unsigned short		port;
 #if IS_ENABLED(CONFIG_IPV6)
+	bool			ipv6_only;
 	unsigned short		addr_type;
 	struct in6_addr		v6_rcv_saddr;
 #define rcv_saddr		v6_rcv_saddr.s6_addr32[3]
diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
index cf88eca5f1b4..5fc749f8f2b1 100644
--- a/net/ipv4/inet_hashtables.c
+++ b/net/ipv4/inet_hashtables.c
@@ -110,6 +110,7 @@ static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb2,
 	tb2->port = tb->port;
 #if IS_ENABLED(CONFIG_IPV6)
 	BUILD_BUG_ON(USHRT_MAX < (IPV6_ADDR_ANY | IPV6_ADDR_MAPPED));
+	tb2->ipv6_only = ipv6_only_sock(sk);
 	if (sk->sk_family == AF_INET6) {
 		tb2->addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
 		tb2->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
@@ -831,7 +832,8 @@ bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const

 #if IS_ENABLED(CONFIG_IPV6)
 	if (tb->addr_type == IPV6_ADDR_ANY)
-		return true;
+		if (sk->sk_family == AF_INET6 || !tb->ipv6_only)
+			return true;

 	if (tb->addr_type != IPV6_ADDR_MAPPED)
 		return false;

> 
> 
>>
>>  	if (tb->addr_type != IPV6_ADDR_MAPPED)
>>  		return false;

Re: [PATCH] tcp: Fix inet_bind2_bucket_match_addr_any() regression

[Kuniyuki Iwashima]

From: Jianguo Wu <wujianguo106@163.com>
Date: Fri, 22 Mar 2024 11:16:09 +0800
> Hi Kuniyuki,
> Thanks for your reply!
> 
> On 2024/3/21 12:55, Kuniyuki Iwashima wrote:
> > Hi,
> > 
> > Thanks for the patch.
> > 
> > From: Jianguo Wu <wujianguo106@163.com>
> > Date: Thu, 21 Mar 2024 11:02:36 +0800
> >> From: Jianguo Wu <wujianguo@chinatelecom.cn>
> >>
> >> If we bind() a TCPv4 socket to 0.0.0.0:8090, then bind() a TCPv6(ipv6only) socket
> > 
> > Please wrap each line at <75 characters except for logs/output like below.
> > 
> OK.
> > 
> >> to :::8090, both without SO_REUSEPORT, then bind() 127.0.0.1:8090, it should fail
> > 
> > [::]:8090 is easier to read and the recommended way.
> > https://datatracker.ietf.org/doc/html/rfc5952#section-6
> > 
> > But please keep the netstat output as is.
> > 
> >> but now succeeds. like this:
> >>   tcp        0      0 127.0.0.1:8090          0.0.0.0:*               LISTEN
> >>   tcp        0      0 0.0.0.0:8090            0.0.0.0:*               LISTEN
> >>   tcp6       0      0 :::8090                 :::*                    LISTEN
> >>
> >> bind() 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.
> > 
> > What do you mean by all fail ?
> > At least, [::1]:8090 would fail with the current code in this case.
> In my test, 127.0.0.1:8090  succeeds, 0.0.0.0:8090, [::]:8090 and [::1]:8090 are all fail
> 
> > 
> > 
> >> But if we bind() a TCPv6(ipv6only) socket to :::8090 first, then  bind() a TCPv4
> >> socket to 0.0.0.0:8090, then bind() 127.0.0.1:8090, 0.0.0.0:8090, :::8090 and ::1:8090 are all fail.
> >>
> >> When bind() 127.0.0.1:8090, inet_bind2_bucket_match_addr_any() will return true as tb->addr_type == IPV6_ADDR_ANY,
> > 
> > Let's use tb2 here for inet_bind2_bucket.. yes it's not consistent
> > in some functions like inet_bind2_bucket_match_addr_any() though.
> yes, inet_bind2_bucket_match_addr_any() use tb, so I use tb here.
> 
> > 
> > 
> >> and tb is refer to the TCPv6 socket(:::8090), then inet_bhash2_conflict() return false, That is, there is no conflict,
> > 
> > Also make it clear that the TCPv6 socket is ipv6only one.
> > 
> > 
> >> so bind() succeeds.
> >>
> >>   inet_bhash2_addr_any_conflict()
> >>   {
> >> 	inet_bind_bucket_for_each(tb2, &head2->chain)
> >> 		// tb2 is IPv6
> >> 		if (inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
> >> 			break;
> >>
> >> 	// inet_bhash2_conflict() return false
> >> 	if (tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
> >>                                 reuseport_ok)) {
> >> 		spin_unlock(&head2->lock);
> >> 		return true;
> >> 	}
> >>
> >>   }
> >>
> >> Fixes: 5a22bba13d01 ("tcp: Save address type in inet_bind2_bucket.")
> > 
> > This is not the commit that introduced the regression.
> I will remove this.
> > 
> > Also, you need Signed-off-by tag here.
> OK.
> > 
> > 
> >> ---
> >>  net/ipv4/inet_hashtables.c | 4 ++--
> >>  1 file changed, 2 insertions(+), 2 deletions(-)
> >>
> >> diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
> >> index 7498af320164..3eeaca8a113f 100644
> >> --- a/net/ipv4/inet_hashtables.c
> >> +++ b/net/ipv4/inet_hashtables.c
> >> @@ -830,8 +830,8 @@ bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const
> >>  		return false;
> >>
> >>  #if IS_ENABLED(CONFIG_IPV6)
> >> -	if (tb->addr_type == IPV6_ADDR_ANY)
> >> -		return true;
> >> +	if (sk->sk_family == AF_INET6)
> >> +		return tb->addr_type == IPV6_ADDR_ANY;
> > 
> > This fix is not correct and will break v4-mapped-v6 address cases.
> > You can run bind_wildcard under the selftest directory.
> >> Probably we need v6_only bit in tb2 and should add some test cases
> > in the selftest.
> How about this?
> I add a new field ipv6_only to struct inet_bind2_bucket{}
> 
> diff --git a/include/net/inet_hashtables.h b/include/net/inet_hashtables.h
> index 7f1b38458743..fb7c250a663b 100644
> --- a/include/net/inet_hashtables.h
> +++ b/include/net/inet_hashtables.h
> @@ -96,6 +96,7 @@ struct inet_bind2_bucket {
>  	int			l3mdev;
>  	unsigned short		port;
>  #if IS_ENABLED(CONFIG_IPV6)
> +	bool			ipv6_only;
>  	unsigned short		addr_type;
>  	struct in6_addr		v6_rcv_saddr;
>  #define rcv_saddr		v6_rcv_saddr.s6_addr32[3]
> diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
> index cf88eca5f1b4..5fc749f8f2b1 100644
> --- a/net/ipv4/inet_hashtables.c
> +++ b/net/ipv4/inet_hashtables.c
> @@ -110,6 +110,7 @@ static void inet_bind2_bucket_init(struct inet_bind2_bucket *tb2,
>  	tb2->port = tb->port;
>  #if IS_ENABLED(CONFIG_IPV6)
>  	BUILD_BUG_ON(USHRT_MAX < (IPV6_ADDR_ANY | IPV6_ADDR_MAPPED));
> +	tb2->ipv6_only = ipv6_only_sock(sk);
>  	if (sk->sk_family == AF_INET6) {
>  		tb2->addr_type = ipv6_addr_type(&sk->sk_v6_rcv_saddr);
>  		tb2->v6_rcv_saddr = sk->sk_v6_rcv_saddr;
> @@ -831,7 +832,8 @@ bool inet_bind2_bucket_match_addr_any(const struct inet_bind2_bucket *tb, const
> 
>  #if IS_ENABLED(CONFIG_IPV6)
>  	if (tb->addr_type == IPV6_ADDR_ANY)
> -		return true;
> +		if (sk->sk_family == AF_INET6 || !tb->ipv6_only)
> +			return true;
> 
>  	if (tb->addr_type != IPV6_ADDR_MAPPED)
>  		return false;
> 

I found a corner case that does not work with this change.
The 3rd bind() should fail but succeeds with the v6only flag approach.

---8<---
from socket import *

s1 = socket(AF_INET6)
s1.setsockopt(41, IPV6_V6ONLY, 1)
s1.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
s1.bind(('::', 8000))

s2 = socket(AF_INET6)
s2.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
s2.bind(('::', 8000))

s3 = socket(AF_INET)
s3.bind(('127.0.0.1', 8000))
---8<---


To cover this case, the only solution is to iterate over the buckets
in the same hash.  Also, this change can be backported to 6.1.

---8<---
diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 612aa1d2eff7..63e8f2df0681 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -288,6 +288,7 @@ static bool inet_bhash2_addr_any_conflict(const struct sock *sk, int port, int l
 	struct sock_reuseport *reuseport_cb;
 	struct inet_bind_hashbucket *head2;
 	struct inet_bind2_bucket *tb2;
+	bool conflict = false;
 	bool reuseport_cb_ok;
 
 	rcu_read_lock();
@@ -300,18 +301,20 @@ static bool inet_bhash2_addr_any_conflict(const struct sock *sk, int port, int l
 
 	spin_lock(&head2->lock);
 
-	inet_bind_bucket_for_each(tb2, &head2->chain)
-		if (inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
-			break;
+	inet_bind_bucket_for_each(tb2, &head2->chain) {
+		if (!inet_bind2_bucket_match_addr_any(tb2, net, port, l3mdev, sk))
+			continue;
 
-	if (tb2 && inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,
-					reuseport_ok)) {
-		spin_unlock(&head2->lock);
-		return true;
+		if (!inet_bhash2_conflict(sk, tb2, uid, relax, reuseport_cb_ok,	reuseport_ok))
+			continue;
+
+		conflict = true;
+		break;
 	}
 
 	spin_unlock(&head2->lock);
-	return false;
+
+	return conflict;
 }
 
 /*
---8<---


Also, I found another regression that my recent patch introduced.
The second bind() should succeed but now fails.

---8<---
from socket import *

s1 = socket(AF_INET6)
s1.setsockopt(41, IPV6_V6ONLY, 1)
s1.bind(('::', 8000))

s2 = socket(AF_INET6)
s2.bind(('::ffff:127.0.0.1', 8000))
---8<---

and we need this change.

---8<---
diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
index 7d8090f109ef..612aa1d2eff7 100644
--- a/net/ipv4/inet_connection_sock.c
+++ b/net/ipv4/inet_connection_sock.c
@@ -203,7 +203,8 @@ static bool __inet_bhash2_conflict(const struct sock *sk, struct sock *sk2,
 				   kuid_t sk_uid, bool relax,
 				   bool reuseport_cb_ok, bool reuseport_ok)
 {
-	if (sk->sk_family == AF_INET && ipv6_only_sock(sk2))
+	if (ipv6_only_sock(sk2) &&
+	    (sk->sk_family == AF_INET || ipv6_addr_v4mapped(&sk->sk_v6_rcv_saddr)))
 		return false;
 
 	return inet_bind_conflict(sk, sk2, sk_uid, relax,
---8<---


I'll post a series including necessary tests.

Thanks!