[patch 0/1][NETNS49] Make af_unix autobind per namespace

[patch 0/1][NETNS49] Make af_unix autobind per namespace

[Daniel Lezcano]

The following patch change autobind fonction to use the ordernum
from the network namespace instead of using the local static variable.

-- 

[patch 1/1][NETNS49] Make af_unix autobind per network namespace

[Daniel Lezcano]

[-- Attachment #1: net-ns-make-af-unix-autobind-per-namespace.patch --]
[-- Type: text/plain, Size: 1664 bytes --]

This patch change the static ordernum variable to be relative to the
network namespace.

Signed-off-by: Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org>
---
 include/net/net_namespace.h |    1 +
 net/unix/af_unix.c          |    5 ++++-
 2 files changed, 5 insertions(+), 1 deletion(-)

Index: linux-2.6-netns/include/net/net_namespace.h
===================================================================
--- linux-2.6-netns.orig/include/net/net_namespace.h
+++ linux-2.6-netns/include/net/net_namespace.h
@@ -60,6 +60,7 @@ struct net {
 	/* AF_UNIX */
 	int			sysctl_unix_max_dgram_qlen;
 	void			*unix_sysctl;
+	u32                     ordernum;
 
 	/* XFRM */
 	u32			sysctl_xfrm_aevent_etime;
Index: linux-2.6-netns/net/unix/af_unix.c
===================================================================
--- linux-2.6-netns.orig/net/unix/af_unix.c
+++ linux-2.6-netns/net/unix/af_unix.c
@@ -681,8 +681,8 @@ static int unix_autobind(struct socket *
 	struct sock *sk = sock->sk;
 	struct net *net = sk->sk_net;
 	struct unix_sock *u = unix_sk(sk);
-	static u32 ordernum = 1;
 	struct unix_address * addr;
+	u32 ordernum = net->ordernum;
 	int err;
 
 	mutex_lock(&u->readlock);
@@ -720,6 +720,7 @@ retry:
 	u->addr = addr;
 	__unix_insert_socket(&unix_socket_table[addr->hash], sk);
 	spin_unlock(&unix_table_lock);
+	net->ordernum = ordernum;
 	err = 0;
 
 out:	mutex_unlock(&u->readlock);
@@ -2164,6 +2165,8 @@ static int unix_net_init(struct net *net
 	int error = -ENOMEM;
 
 	net->sysctl_unix_max_dgram_qlen = 10;
+	net->ordernum = 1;
+
 #ifdef CONFIG_PROC_FS
 	if (!proc_net_fops_create(net, "unix", 0, &unix_seq_fops))
 		goto out;

-- 

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Daniel Lezcano]

Daniel Lezcano wrote:
> The following patch change autobind fonction to use the ordernum
> from the network namespace instead of using the local static variable.
> 

I forgot to add this trivial program which does autobind.

---


#include <sys/socket.h>
#include <sys/un.h>
#include <unistd.h>
#include <errno.h>
#include <string.h>
#include <stdio.h>

int main(int argc, char* argv[])
{
         int fd;
         struct sockaddr_un addr;

         fd = socket(PF_UNIX, SOCK_DGRAM, 0);
         if (fd == -1) {
                 perror("socket");
                 return 1;
         }

         memset(&addr, 0, sizeof(addr));

         addr.sun_family = AF_UNIX;
         strcpy(addr.sun_path, "");

         if (bind(fd, &addr, sizeof(short))) {
                 perror("bind");
                 return 1;
         }

         return 0;
}

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Eric W. Biederman]

Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:

> The following patch change autobind fonction to use the ordernum
> from the network namespace instead of using the local static variable.

Why do we care?
Information leak?
Some application is expecting a predictable autobind value?

Just skimming the code it looks like it will work correctly without
this.

Eric

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Daniel Lezcano]

Eric W. Biederman wrote:
> Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
> 
>> The following patch change autobind fonction to use the ordernum
>> from the network namespace instead of using the local static variable.
> 
> Why do we care?
> Information leak?
> Some application is expecting a predictable autobind value?
> 
> Just skimming the code it looks like it will work correctly without
> this.

I think my summary is ... too short :)

I don't see any applications taking care of this. If they ask for an 
abstract socket, then they don't care about the bind result. So 
probably, the patchset is totally useless.

But from the POV of the checkpoint/restart, we should check if this 
value is somewhere visible from userspace and so storable by an application.

It appears this is the case with /proc/net/unix, where an abstract 
socket is symbolized by the path pattern "@". Example:

cat /proc/net/unix

Num       RefCount Protocol Flags    Type St Inode Path
c6a27710: 00000002 00000000 00000000 0002 01  4357 @00003

I agree by the fact that can be considered as a detail and the 
probability to have an application storing this informaton is very small 
( eg. checkpointing while doing netstat in the container ). But IMHO, 
the paradigm "never seen from userspace" fails and that justifies to 
have the ordernum variable relative to a namespace.

   -- Daniel

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Eric W. Biederman]

Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:

> Eric W. Biederman wrote:
>> Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
>>
>>> The following patch change autobind fonction to use the ordernum
>>> from the network namespace instead of using the local static variable.
>>
>> Why do we care?
>> Information leak?
>> Some application is expecting a predictable autobind value?
>>
>> Just skimming the code it looks like it will work correctly without
>> this.
>
> I think my summary is ... too short :)
>
> I don't see any applications taking care of this. If they ask for an abstract
> socket, then they don't care about the bind result. So probably, the patchset is
> totally useless.
>
> But from the POV of the checkpoint/restart, we should check if this value is
> somewhere visible from userspace and so storable by an application.

Right.  And we already can already specifically select this result.
My point is that the semi random sequence generator logic does not
need to be per namespace, because people don't care what the sequence.
That sequence is not exported to user space.

> It appears this is the case with /proc/net/unix, where an abstract socket is
> symbolized by the path pattern "@". Example:
>
> cat /proc/net/unix
>
> Num       RefCount Protocol Flags    Type St Inode Path
> c6a27710: 00000002 00000000 00000000 0002 01  4357 @00003

Right, and that part we should definitely preserve for checkpoint/restart
purposes.

> I agree by the fact that can be considered as a detail and the probability to
> have an application storing this informaton is very small ( eg. checkpointing
> while doing netstat in the container ). But IMHO, the paradigm "never seen from
> userspace" fails and that justifies to have the ordernum variable relative to a
> namespace.

My point was that ordernum itself is not seen.  It is just an arbitrary number
and we are allowed to change the algorithm for selecting a new abstract
namespace name at will.

If there is something in userspace that depends on the algorithm for selecting
the abstract name then making ordernum per namespace make sense.

At the moment the code is simpler to use what is effectively a different
algorithm for selecting the abstract namespace name of the socket.

Eric

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Daniel Lezcano]

Eric W. Biederman wrote:
> Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
> 
>> Eric W. Biederman wrote:
>>> Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
>>>
>>>> The following patch change autobind fonction to use the ordernum
>>>> from the network namespace instead of using the local static variable.
>>> Why do we care?
>>> Information leak?
>>> Some application is expecting a predictable autobind value?
>>>
>>> Just skimming the code it looks like it will work correctly without
>>> this.
>> I think my summary is ... too short :)
>>
>> I don't see any applications taking care of this. If they ask for an abstract
>> socket, then they don't care about the bind result. So probably, the patchset is
>> totally useless.
>>
>> But from the POV of the checkpoint/restart, we should check if this value is
>> somewhere visible from userspace and so storable by an application.
> 
> Right.  And we already can already specifically select this result.
> My point is that the semi random sequence generator logic does not
> need to be per namespace, because people don't care what the sequence.
> That sequence is not exported to user space.
> 
>> It appears this is the case with /proc/net/unix, where an abstract socket is
>> symbolized by the path pattern "@". Example:
>>
>> cat /proc/net/unix
>>
>> Num       RefCount Protocol Flags    Type St Inode Path
>> c6a27710: 00000002 00000000 00000000 0002 01  4357 @00003
> 
> Right, and that part we should definitely preserve for checkpoint/restart
> purposes.
> 
>> I agree by the fact that can be considered as a detail and the probability to
>> have an application storing this informaton is very small ( eg. checkpointing
>> while doing netstat in the container ). But IMHO, the paradigm "never seen from
>> userspace" fails and that justifies to have the ordernum variable relative to a
>> namespace.
> 
> My point was that ordernum itself is not seen.  It is just an arbitrary number
> and we are allowed to change the algorithm for selecting a new abstract
> namespace name at will.

Hmm, right. That makes sense.

> If there is something in userspace that depends on the algorithm for selecting
> the abstract name then making ordernum per namespace make sense.

Ok, fair enough. Let's forget this patch. It is small enough to rewrite 
it if unexpectedly something bad happens with ordernum.

  -- Daniel

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Denis V. Lunev]

Daniel Lezcano wrote:
> Eric W. Biederman wrote:
>> Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
>>
>>> The following patch change autobind fonction to use the ordernum
>>> from the network namespace instead of using the local static variable.
>>
>> Why do we care?
>> Information leak?
>> Some application is expecting a predictable autobind value?
>>
>> Just skimming the code it looks like it will work correctly without
>> this.
> 
> I think my summary is ... too short :)
> 
> I don't see any applications taking care of this. If they ask for an
> abstract socket, then they don't care about the bind result. So
> probably, the patchset is totally useless.
> 
> But from the POV of the checkpoint/restart, we should check if this
> value is somewhere visible from userspace and so storable by an
> application.

we do not care with this in checkpointing. One namespace socket does not
see other namespace socket

Regards,
	Den

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Cedric Le Goater]

Denis V. Lunev wrote:
> Daniel Lezcano wrote:
>> Eric W. Biederman wrote:
>>> Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
>>>
>>>> The following patch change autobind fonction to use the ordernum
>>>> from the network namespace instead of using the local static variable.
>>> Why do we care?
>>> Information leak?
>>> Some application is expecting a predictable autobind value?
>>>
>>> Just skimming the code it looks like it will work correctly without
>>> this.
>> I think my summary is ... too short :)
>>
>> I don't see any applications taking care of this. If they ask for an
>> abstract socket, then they don't care about the bind result. So
>> probably, the patchset is totally useless.
>>
>> But from the POV of the checkpoint/restart, we should check if this
>> value is somewhere visible from userspace and so storable by an
>> application.
> 
> we do not care with this in checkpointing. One namespace socket does not
> see other namespace socket

my 2 cnts,

when 'restarting' a socket bound to an abstract name, we will have 
a EADDRINUSE if we try to rebind it to an abtract name which is 
already in use by a socket in a another namespace ? 

it seems to me that this is an identifier and like any identifier
it should be private to the namespace, which probably means having
unix_abstract_socket_table[] per net namespace.

Cheers,

C. 

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Eric W. Biederman]

Cedric Le Goater <clg-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
> my 2 cnts,
>
> when 'restarting' a socket bound to an abstract name, we will have 
> a EADDRINUSE if we try to rebind it to an abtract name which is 
> already in use by a socket in a another namespace ? 

No.

> it seems to me that this is an identifier and like any identifier
> it should be private to the namespace, which probably means having
> unix_abstract_socket_table[] per net namespace.

Yes it is.  It is a hash table so we are filter the hash chain
and not having two copies of the table.  But effectively it's
the same thing.

All this patch was suggesting was having a per network namespace
copy of the data structure for the random number generator for
generating the name.  

The ``random number generator'' is just a 16bit counter that loops
through all 64k values seeing if a name is available and if so
using it.  Sharing our place in the loop between different namespaces
may be ineffeicient but it should work fine.

Eric

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Cedric Le Goater]

Eric W. Biederman wrote:
> Cedric Le Goater <clg-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
>> my 2 cnts,
>>
>> when 'restarting' a socket bound to an abstract name, we will have 
>> a EADDRINUSE if we try to rebind it to an abtract name which is 
>> already in use by a socket in a another namespace ? 
> 
> No.

ok. i just went over your AF_UNIX patch and saw that 'struct net' was 
being checked for abstract sockets.

C.

>> it seems to me that this is an identifier and like any identifier
>> it should be private to the namespace, which probably means having
>> unix_abstract_socket_table[] per net namespace.
> 
> Yes it is.  It is a hash table so we are filter the hash chain
> and not having two copies of the table.  But effectively it's
> the same thing.
> 
> All this patch was suggesting was having a per network namespace
> copy of the data structure for the random number generator for
> generating the name.  
> 
> The ``random number generator'' is just a 16bit counter that loops
> through all 64k values seeing if a name is available and if so
> using it.  Sharing our place in the loop between different namespaces
> may be ineffeicient but it should work fine.
> 
> Eric

Re: [patch 0/1][NETNS49] Make af_unix autobind per namespace

[Denis V. Lunev]

Eric W. Biederman wrote:
> Daniel Lezcano <dlezcano-NmTC/0ZBporQT0dZR+AlfA@public.gmane.org> writes:
> 
>> The following patch change autobind fonction to use the ordernum
>> from the network namespace instead of using the local static variable.
> 
> Why do we care?
> Information leak?
> Some application is expecting a predictable autobind value?
> 
> Just skimming the code it looks like it will work correctly without
> this.

I also do not see a need for this :)

Regards,
	Den