Re: [PATCH net v2 2/3] vsock: lock down child_ns_mode as write-once

[Stefano Garzarella <sgarzare@redhat.com>]

On Thu, Feb 19, 2026 at 08:20:54AM -0800, Bobby Eshleman wrote:
>On Thu, Feb 19, 2026 at 11:35:52AM +0100, Stefano Garzarella wrote:
>> On Wed, Feb 18, 2026 at 10:10:37AM -0800, Bobby Eshleman wrote:
>> > From: Bobby Eshleman <bobbyeshleman@meta.com>
>> >
>> > Two administrator processes may race when setting child_ns_mode as one
>> > process sets child_ns_mode to "local" and then creates a namespace, but
>> > another process changes child_ns_mode to "global" between the write and
>> > the namespace creation. The first process ends up with a namespace in
>> > "global" mode instead of "local". While this can be detected after the
>> > fact by reading ns_mode and retrying, it is fragile and error-prone.
>> >
>> > Make child_ns_mode write-once so that a namespace manager can set it
>> > once and be sure it won't change. Writing a different value after the
>> > first write returns -EBUSY. This applies to all namespaces, including
>> > init_net, where an init process can write "local" to lock all future
>> > namespaces into local mode.
>> >
>> > Fixes: eafb64f40ca4 ("vsock: add netns to vsock core")
>> > Signed-off-by: Bobby Eshleman <bobbyeshleman@meta.com>
>> > Suggested-by: Daan De Meyer <daan.j.demeyer@gmail.com>
>> > Suggested-by: Stefano Garzarella <sgarzare@redhat.com>
>>
>> nit: usually the S-o-b of the author is the last when sending a patch.
>
>Ah good to know, thanks. Will change.
>
>>
>> > ---
>> > include/net/af_vsock.h    | 20 +++++++++++++++++---
>> > include/net/netns/vsock.h |  9 ++++++++-
>> > net/vmw_vsock/af_vsock.c  | 15 ++++++++++-----
>> > 3 files changed, 35 insertions(+), 9 deletions(-)
>> >
>> > diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>> > index d3ff48a2fbe0..9bd42147626d 100644
>> > --- a/include/net/af_vsock.h
>> > +++ b/include/net/af_vsock.h
>> > @@ -276,15 +276,29 @@ static inline bool vsock_net_mode_global(struct vsock_sock *vsk)
>> > 	return vsock_net_mode(sock_net(sk_vsock(vsk))) == VSOCK_NET_MODE_GLOBAL;
>> > }
>> >
>> > -static inline void vsock_net_set_child_mode(struct net *net,
>> > +static inline bool vsock_net_set_child_mode(struct net *net,
>> > 					    enum vsock_net_mode mode)
>> > {
>> > -	WRITE_ONCE(net->vsock.child_ns_mode, mode);
>> > +	int locked = mode + VSOCK_NET_MODE_LOCKED;
>> > +	int cur;
>> > +
>> > +	cur = READ_ONCE(net->vsock.child_ns_mode);
>> > +	if (cur == locked)
>> > +		return true;
>> > +	if (cur >= VSOCK_NET_MODE_LOCKED)
>> > +		return false;
>> > +
>> > +	if (try_cmpxchg(&net->vsock.child_ns_mode, &cur, locked))
>> > +		return true;
>> > +
>> > +	return cur == locked;
>>
>> Sorry, it took me a while to get it entirely :-(
>> This overcomplication is exactly what I wanted to avoid when I proposed the
>> change in v1:
>> https://lore.kernel.org/netdev/aZWUmbiH11Eh3Y4v@sgarzare-redhat/
>
>Glad you thought so too, because I actually think your original proposed
>snippet in that thread is the best/simplest so far.
>
>>
>>
>> > }
>> >
>> > static inline enum vsock_net_mode vsock_net_child_mode(struct net *net)
>> > {
>> > -	return READ_ONCE(net->vsock.child_ns_mode);
>> > +	int mode = READ_ONCE(net->vsock.child_ns_mode);
>> > +
>> > +	return mode & (VSOCK_NET_MODE_LOCKED - 1);
>>
>> This is working just because VSOCK_NET_MODE_LOCKED == 2, so IMO this should
>> at least set as value in the enum and documented on top of vsock_net_mode.
>>
>> > }
>> >
>> > /* Return true if two namespaces pass the mode rules. Otherwise, return false.
>> > diff --git a/include/net/netns/vsock.h b/include/net/netns/vsock.h
>> > index b34d69a22fa8..d20ab6269342 100644
>> > --- a/include/net/netns/vsock.h
>> > +++ b/include/net/netns/vsock.h
>> > @@ -7,6 +7,7 @@
>> > enum vsock_net_mode {
>> > 	VSOCK_NET_MODE_GLOBAL,
>> > 	VSOCK_NET_MODE_LOCAL,
>> > +	VSOCK_NET_MODE_LOCKED,
>>
>> This is not really a mode, so IMO should not be part of `enum
>> vsock_net_mode`. If you really want it, maybe we can add both
>> VSOCK_NET_MODE_GLOBAL_LOCKED and VSOCK_NET_MODE_LOCAL_LOCKED, which can be
>> less error prone if we will touch this enum one day.
>>
>> > };
>> >
>> > struct netns_vsock {
>> > @@ -16,6 +17,12 @@ struct netns_vsock {
>> > 	u32 port;
>> >
>> > 	enum vsock_net_mode mode;
>> > -	enum vsock_net_mode child_ns_mode;
>> > +
>> > +	/* 0 (GLOBAL)
>> > +	 * 1 (LOCAL)
>> > +	 * 2 (GLOBAL + LOCKED)
>> > +	 * 3 (LOCAL + LOCKED)
>> > +	 */
>> > +	int child_ns_mode;
>>
>> Sorry, I don't like this too much, since it seems too complicated to read
>> and to maintain, If we really want to use just one variable, maybe we can
>> use -1 as UNSET for child_ns_mode. If it is UNSET, vsock_net_child_mode()
>> can just return `mode` since it's the default that we also documented, if
>> it's set, it means that is locked with the value specified.
>>
>> Maybe with code is easier, I mean something like this:
>>
>> diff --git a/include/net/af_vsock.h b/include/net/af_vsock.h
>> index d3ff48a2fbe0..fcd5b538df35 100644
>> --- a/include/net/af_vsock.h
>> +++ b/include/net/af_vsock.h
>> @@ -276,15 +276,25 @@ static inline bool vsock_net_mode_global(struct vsock_sock *vsk)
>>  	return vsock_net_mode(sock_net(sk_vsock(vsk))) == VSOCK_NET_MODE_GLOBAL;
>>  }
>> -static inline void vsock_net_set_child_mode(struct net *net,
>> +static inline bool vsock_net_set_child_mode(struct net *net,
>>  					    enum vsock_net_mode mode)
>>  {
>> -	WRITE_ONCE(net->vsock.child_ns_mode, mode);
>> +	int old = VSOCK_NET_CHILD_NS_UNSET;
>> +
>> +	if (try_cmpxchg(&net->vsock.child_ns_mode, &old, mode))
>> +		return true;
>> +
>> +	return old == mode;
>>  }
>>  static inline enum vsock_net_mode vsock_net_child_mode(struct net *net)
>>  {
>> -	return READ_ONCE(net->vsock.child_ns_mode);
>> +	int mode = READ_ONCE(net->vsock.child_ns_mode);
>> +
>> +	if (mode == VSOCK_NET_CHILD_NS_UNSET)
>> +		return net->vsock.mode;
>> +
>> +	return mode;
>>  }
>>  /* Return true if two namespaces pass the mode rules. Otherwise, return false.
>> diff --git a/include/net/netns/vsock.h b/include/net/netns/vsock.h
>> index b34d69a22fa8..bf52baf7d7a7 100644
>> --- a/include/net/netns/vsock.h
>> +++ b/include/net/netns/vsock.h
>> @@ -9,6 +9,8 @@ enum vsock_net_mode {
>>  	VSOCK_NET_MODE_LOCAL,
>>  };
>> +#define VSOCK_NET_CHILD_NS_UNSET (-1)
>> +
>>  struct netns_vsock {
>>  	struct ctl_table_header *sysctl_hdr;
>> @@ -16,6 +18,13 @@ struct netns_vsock {
>>  	u32 port;
>>  	enum vsock_net_mode mode;
>> -	enum vsock_net_mode child_ns_mode;
>> +
>> +	/* Write-once child namespace mode, must be initialized to
>> +	 * VSOCK_NET_CHILD_NS_UNSET. Transitions once from UNSET to a
>> +	 * vsock_net_mode value via try_cmpxchg on first sysctl write.
>> +	 * While UNSET, vsock_net_child_mode() returns the namespace's
>> +	 * own mode since it's the default.
>> +	 */
>> +	int child_ns_mode;
>>  };
>>  #endif /* __NET_NET_NAMESPACE_VSOCK_H */
>> diff --git a/net/vmw_vsock/af_vsock.c b/net/vmw_vsock/af_vsock.c
>> index 9880756d9eff..f0cb7c6a8212 100644
>> --- a/net/vmw_vsock/af_vsock.c
>> +++ b/net/vmw_vsock/af_vsock.c
>> @@ -2853,7 +2853,8 @@ static int vsock_net_child_mode_string(const struct ctl_table *table, int write,
>>  		    new_mode == VSOCK_NET_MODE_GLOBAL)
>>  			return -EPERM;
>> -		vsock_net_set_child_mode(net, new_mode);
>> +		if (!vsock_net_set_child_mode(net, new_mode))
>> +			return -EBUSY;
>>  	}
>>  	return 0;
>> @@ -2922,7 +2923,7 @@ static void vsock_net_init(struct net *net)
>>  	else
>>  		net->vsock.mode = vsock_net_child_mode(current->nsproxy->net_ns);
>> -	net->vsock.child_ns_mode = net->vsock.mode;
>> +	net->vsock.child_ns_mode = VSOCK_NET_CHILD_NS_UNSET;
>>  }
>>  static __net_init int vsock_sysctl_init_net(struct net *net)
>>
>> If you like it, please add my Co-developed-by and S-o-b.
>
>Will do!
>
>>
>> BTW, let's discuss here more about it and agree before sending a new
>> version, so this should also allow other to comment eventually.
>>
>> Thanks,
>> Stefano
>
>Tbh, I like your original proposal from v1 best (copied below). I like
>that the whole locking mechanism is self-contained there in one place,
>and doesn't ripple out elsewhere into the code (e.g.,
>vsock_net_child_mode() carrying logic around UNSET). Wdyt?

Initially, yes, I liked that one too, especially because, being a patch 
for net, it remains very small and clear to read. But now, after 
spending some time on how to reuse `child_ns_mode` for that, I also like 
the last version I sent using UNSET so that we don't have the same 
information in two variables.

I'm truly conflicted, but not a strong preference, so if you like more 
the one with `child_ns_mode_locked`, let's go with that, we can always 
change it in the future.

Jacub, Paolo, any preference?

>
>static inline bool vsock_net_set_child_mode(struct net *net,
>					    enum vsock_net_mode mode)
>{
>	int new_locked = mode + 1;
>	int old_locked = 0;

If we are going to use this one, maybe a macro for 0, or a comment here 
+ on top of child_ns_mode_locked should be better.

Thanks,
Stefano

>
>	if (try_cmpxchg(&net->vsock.child_ns_mode_locked,
>			&old_locked, new_locked)) {
>		WRITE_ONCE(net->vsock.child_ns_mode, mode);
>		return true;
>	}
>
>	return old_locked == new_locked;
>}
>
>
>Best,
>Bobby
>
>>
>> > };
>> > #endif /* __NET_NET_NAMESPACE_VSOCK_H */
>> > diff --git a/net/vmw_vsock/af_vsock.c b/net/vmw_vsock/af_vsock.c
>> > index 9880756d9eff..50044a838c89 100644
>> > --- a/net/vmw_vsock/af_vsock.c
>> > +++ b/net/vmw_vsock/af_vsock.c
>> > @@ -90,16 +90,20 @@
>> >  *
>> >  *   - /proc/sys/net/vsock/ns_mode (read-only) reports the current namespace's
>> >  *     mode, which is set at namespace creation and immutable thereafter.
>> > - *   - /proc/sys/net/vsock/child_ns_mode (writable) controls what mode future
>> > + *   - /proc/sys/net/vsock/child_ns_mode (write-once) controls what mode future
>> >  *     child namespaces will inherit when created. The initial value matches
>> >  *     the namespace's own ns_mode.
>> >  *
>> >  *   Changing child_ns_mode only affects newly created namespaces, not the
>> >  *   current namespace or existing children. A "local" namespace cannot set
>> > - *   child_ns_mode to "global". At namespace creation, ns_mode is inherited
>> > - *   from the parent's child_ns_mode.
>> > + *   child_ns_mode to "global". child_ns_mode is write-once, so that it may be
>> > + *   configured and locked down by a namespace manager. Writing a different
>> > + *   value after the first write returns -EBUSY. At namespace creation, ns_mode
>> > + *   is inherited from the parent's child_ns_mode.
>> >  *
>> > - *   The init_net mode is "global" and cannot be modified.
>> > + *   The init_net mode is "global" and cannot be modified. The init_net
>> > + *   child_ns_mode is also write-once, so an init process (e.g. systemd) can
>> > + *   set it to "local" to ensure all new namespaces inherit local mode.
>> >  *
>> >  *   The modes affect the allocation and accessibility of CIDs as follows:
>> >  *
>> > @@ -2853,7 +2857,8 @@ static int vsock_net_child_mode_string(const struct ctl_table *table, int write,
>> > 		    new_mode == VSOCK_NET_MODE_GLOBAL)
>> > 			return -EPERM;
>> >
>> > -		vsock_net_set_child_mode(net, new_mode);
>> > +		if (!vsock_net_set_child_mode(net, new_mode))
>> > +			return -EBUSY;
>> > 	}
>> >
>> > 	return 0;
>> >
>> > --
>> > 2.47.3
>> >
>>
>