[PATCH ipsec 1/2] xfrm: Wait for RCU readers during policy netns exit

[PATCH ipsec 1/2] xfrm: Wait for RCU readers during policy netns exit

[Steffen Klassert]

xfrm_policy_fini() frees the policy_bydst hash tables after flushing the
policy work items and deleting all policies, but it does not wait for
concurrent RCU readers to leave their read-side critical sections first.

The policy_bydst tables are published via rcu_assign_pointer() and are
looked up through rcu_dereference_check(), so netns teardown must also
wait for an RCU grace period before freeing the table memory.

Fix this by adding synchronize_rcu() before freeing the policy hash tables.

Fixes: e1e551bc5630 ("xfrm: policy: prepare policy_bydst hash for rcu lookups")
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
---
 net/xfrm/xfrm_policy.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/xfrm/xfrm_policy.c b/net/xfrm/xfrm_policy.c
index 362939aa56cf..8f0188e763c7 100644
--- a/net/xfrm/xfrm_policy.c
+++ b/net/xfrm/xfrm_policy.c
@@ -4290,6 +4290,8 @@ static void xfrm_policy_fini(struct net *net)
 #endif
 	xfrm_policy_flush(net, XFRM_POLICY_TYPE_MAIN, false);
 
+	synchronize_rcu();
+
 	WARN_ON(!list_empty(&net->xfrm.policy_all));
 
 	for (dir = 0; dir < XFRM_POLICY_MAX; dir++) {
-- 
2.43.0

[PATCH ipsec 2/2] xfrm: Wait for RCU readers during state netns exit

[Steffen Klassert]

xfrm_state_fini() flushes the resize and GC work, destroys all states, and
then frees the state hash tables and the inbound percpu state cache.
Those objects can still be observed by concurrent RCU readers. We need
to wait for a RCU grace period before freeing the hash tables and the
percpu cache to avoid netns teardown racing with lockless lookups.

Fix this by adding synchronize_rcu() before freeing the state hash
tables and the inbound percpu state cache.

Fixes: c8406998b801 ("xfrm: state: use rcu_deref and assign_pointer helpers")
Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
---
 net/xfrm/xfrm_state.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
index 1748d374abca..84fbf1591138 100644
--- a/net/xfrm/xfrm_state.c
+++ b/net/xfrm/xfrm_state.c
@@ -3327,6 +3327,8 @@ void xfrm_state_fini(struct net *net)
 	xfrm_state_flush(net, 0, false);
 	flush_work(&xfrm_state_gc_work);
 
+	synchronize_rcu();
+
 	WARN_ON(!list_empty(&net->xfrm.state_all));
 
 	for (i = 0; i <= net->xfrm.state_hmask; i++) {
-- 
2.43.0

Re: [PATCH ipsec 2/2] xfrm: Wait for RCU readers during state netns exit

[Florian Westphal]

Steffen Klassert <steffen.klassert@secunet.com> wrote:
> xfrm_state_fini() flushes the resize and GC work, destroys all states, and
> then frees the state hash tables and the inbound percpu state cache.
> Those objects can still be observed by concurrent RCU readers. We need
> to wait for a RCU grace period before freeing the hash tables and the
> percpu cache to avoid netns teardown racing with lockless lookups.
> 
> Fix this by adding synchronize_rcu() before freeing the state hash
> tables and the inbound percpu state cache.
> 
> Fixes: c8406998b801 ("xfrm: state: use rcu_deref and assign_pointer helpers")
> Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
> ---
>  net/xfrm/xfrm_state.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
> index 1748d374abca..84fbf1591138 100644
> --- a/net/xfrm/xfrm_state.c
> +++ b/net/xfrm/xfrm_state.c
> @@ -3327,6 +3327,8 @@ void xfrm_state_fini(struct net *net)
>  	xfrm_state_flush(net, 0, false);
>  	flush_work(&xfrm_state_gc_work);
>  
> +	synchronize_rcu();
> +

xfrm_state_flush() is expected to relink all states to the gc list for
the gc worker to handle.

The xfrm state worker calls synchronize_rcu() and then frees states
on the state gc list.

What scenario does this additional synchronize_rcu() protect against?

AFAICS this one is not needed, but maybe i am missing something?

Re: [PATCH ipsec 2/2] xfrm: Wait for RCU readers during state netns exit

[Steffen Klassert]

On Thu, Apr 02, 2026 at 01:57:37PM +0200, Florian Westphal wrote:
> Steffen Klassert <steffen.klassert@secunet.com> wrote:
> > xfrm_state_fini() flushes the resize and GC work, destroys all states, and
> > then frees the state hash tables and the inbound percpu state cache.
> > Those objects can still be observed by concurrent RCU readers. We need
> > to wait for a RCU grace period before freeing the hash tables and the
> > percpu cache to avoid netns teardown racing with lockless lookups.
> > 
> > Fix this by adding synchronize_rcu() before freeing the state hash
> > tables and the inbound percpu state cache.
> > 
> > Fixes: c8406998b801 ("xfrm: state: use rcu_deref and assign_pointer helpers")
> > Signed-off-by: Steffen Klassert <steffen.klassert@secunet.com>
> > ---
> >  net/xfrm/xfrm_state.c | 2 ++
> >  1 file changed, 2 insertions(+)
> > 
> > diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
> > index 1748d374abca..84fbf1591138 100644
> > --- a/net/xfrm/xfrm_state.c
> > +++ b/net/xfrm/xfrm_state.c
> > @@ -3327,6 +3327,8 @@ void xfrm_state_fini(struct net *net)
> >  	xfrm_state_flush(net, 0, false);
> >  	flush_work(&xfrm_state_gc_work);
> >  
> > +	synchronize_rcu();
> > +
> 
> xfrm_state_flush() is expected to relink all states to the gc list for
> the gc worker to handle.
> 
> The xfrm state worker calls synchronize_rcu() and then frees states
> on the state gc list.

Yes, xfrm_state_gc_task() does the synchronize_rcu() already. I found
the ploicy thing and had the impression we need it here as well.

I'll skip this one out.

Thanks!

Re: [PATCH ipsec 1/2] xfrm: Wait for RCU readers during policy netns exit

[Florian Westphal]

Steffen Klassert <steffen.klassert@secunet.com> wrote:
> xfrm_policy_fini() frees the policy_bydst hash tables after flushing the
> policy work items and deleting all policies, but it does not wait for
> concurrent RCU readers to leave their read-side critical sections first.
> 
> The policy_bydst tables are published via rcu_assign_pointer() and are
> looked up through rcu_dereference_check(), so netns teardown must also
> wait for an RCU grace period before freeing the table memory.
> 
> Fix this by adding synchronize_rcu() before freeing the policy hash tables.
> 
> Fixes: e1e551bc5630 ("xfrm: policy: prepare policy_bydst hash for rcu lookups")

Reviewed-by: Florian Westphal <fw@strlen.de>