[PATCH v3 net-next 00/10] rtnetlink: Convert rtnl_newlink() to per-netns RTNL.

[PATCH v3 net-next 00/10] rtnetlink: Convert rtnl_newlink() to per-netns RTNL.

[Kuniyuki Iwashima]

Patch 1 - 3 removes __rtnl_link_unregister() and protects link_ops by
its dedicated mutex to move synchronize_srcu() out of the RTNL scope.

Patch 4 introduces struct rtnl_nets and helper functions to acquire
multiple per-netns RTNL in rtnl_newlink().

Patch 5 - 8 are to prefetch the peer device's netns in rtnl_newlink().

Patch 9 converts rtnl_newlink() to per-netns RTNL.

Patch 10 pushes RTNL down to rtnl_dellink() and rtnl_setlink(), but
the conversion will not be completed unless we support cases with
peer/upper/lower devices.

I confirmed v3 survived ./rtnetlink.sh; rmmod netdevsim.ko; without
lockdep splat.


Changes:
  v3:
    * Add patch 1~3 to avoid SRCU & RTNL deadlock

  v2: https://lore.kernel.org/netdev/20241106022432.13065-1-kuniyu@amazon.com/
    * Patch 1
      * Move struct rtnl_nets to rtnetlink.c
      * Unexport rtnl_nets_add()
    * Patch 2
      * Rename the helper to rtnl_link_get_net_ifla()
      * Unexport rtnl_link_get_net_ifla()
      * Change peer_type to u16
    * Patch 6
      * Remove __rtnl_unlock() dance

  v1: https://lore.kernel.org/netdev/20241105020514.41963-1-kuniyu@amazon.com/


Kuniyuki Iwashima (10):
  rtnetlink: Remove __rtnl_link_unregister().
  rtnetlink: Protect link_ops by mutex.
  rtnetlink: Remove __rtnl_link_register()
  rtnetlink: Introduce struct rtnl_nets and helpers.
  rtnetlink: Add peer_type in struct rtnl_link_ops.
  veth: Set VETH_INFO_PEER to veth_link_ops.peer_type.
  vxcan: Set VXCAN_INFO_PEER to vxcan_link_ops.peer_type.
  netkit: Set IFLA_NETKIT_PEER_INFO to netkit_link_ops.peer_type.
  rtnetlink: Convert RTM_NEWLINK to per-netns RTNL.
  rtnetlink: Register rtnl_dellink() and rtnl_setlink() with
    RTNL_FLAG_DOIT_PERNET_WIP.

 drivers/net/can/vxcan.c  |  12 +-
 drivers/net/dummy.c      |  17 ++-
 drivers/net/ifb.c        |  17 ++-
 drivers/net/netkit.c     |  11 +-
 drivers/net/veth.c       |  18 +--
 include/net/rtnetlink.h  |   8 +-
 net/core/net_namespace.c |   1 -
 net/core/rtnetlink.c     | 257 ++++++++++++++++++++++++++++-----------
 8 files changed, 217 insertions(+), 124 deletions(-)

-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 01/10] rtnetlink: Remove __rtnl_link_unregister().

[Kuniyuki Iwashima]

rtnl_link_unregister() holds RTNL and calls __rtnl_link_unregister(),
where we call synchronize_srcu() to wait inflight RTM_NEWLINK requests
for per-netns RTNL.

We put synchronize_srcu() in __rtnl_link_unregister() due to ifb.ko
and dummy.ko.

However, rtnl_newlink() will acquire SRCU before RTNL later in this
series.  Then, lockdep will detect the deadlock:

   rtnl_link_unregister()       rtnl_newlink()
   ----                         ----
   lock(rtnl_mutex);
                                lock(&ops->srcu);
                                lock(rtnl_mutex);
   sync(&ops->srcu);

To avoid the problem, we must call synchronize_srcu() before RTNL in
rtnl_link_unregister().

As a preparation, let's remove __rtnl_link_unregister().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 drivers/net/dummy.c     |  6 +++++-
 drivers/net/ifb.c       |  6 +++++-
 include/net/rtnetlink.h |  1 -
 net/core/rtnetlink.c    | 32 ++++++++++----------------------
 4 files changed, 20 insertions(+), 25 deletions(-)

diff --git a/drivers/net/dummy.c b/drivers/net/dummy.c
index e9c5e1e11fa0..72618b6af44e 100644
--- a/drivers/net/dummy.c
+++ b/drivers/net/dummy.c
@@ -166,6 +166,7 @@ static int __init dummy_init_one(void)
 
 static int __init dummy_init_module(void)
 {
+	bool need_unregister = false;
 	int i, err = 0;
 
 	down_write(&pernet_ops_rwsem);
@@ -179,12 +180,15 @@ static int __init dummy_init_module(void)
 		cond_resched();
 	}
 	if (err < 0)
-		__rtnl_link_unregister(&dummy_link_ops);
+		need_unregister = true;
 
 out:
 	rtnl_unlock();
 	up_write(&pernet_ops_rwsem);
 
+	if (need_unregister)
+		rtnl_link_unregister(&dummy_link_ops);
+
 	return err;
 }
 
diff --git a/drivers/net/ifb.c b/drivers/net/ifb.c
index 2c1b5def4a0b..a4b9ec4e8f30 100644
--- a/drivers/net/ifb.c
+++ b/drivers/net/ifb.c
@@ -424,6 +424,7 @@ static int __init ifb_init_one(int index)
 
 static int __init ifb_init_module(void)
 {
+	bool need_unregister = false;
 	int i, err;
 
 	down_write(&pernet_ops_rwsem);
@@ -437,12 +438,15 @@ static int __init ifb_init_module(void)
 		cond_resched();
 	}
 	if (err)
-		__rtnl_link_unregister(&ifb_link_ops);
+		need_unregister = true;
 
 out:
 	rtnl_unlock();
 	up_write(&pernet_ops_rwsem);
 
+	if (need_unregister)
+		rtnl_link_unregister(&ifb_link_ops);
+
 	return err;
 }
 
diff --git a/include/net/rtnetlink.h b/include/net/rtnetlink.h
index b260c0cc9671..3ebfcc6e56fd 100644
--- a/include/net/rtnetlink.h
+++ b/include/net/rtnetlink.h
@@ -165,7 +165,6 @@ struct rtnl_link_ops {
 };
 
 int __rtnl_link_register(struct rtnl_link_ops *ops);
-void __rtnl_link_unregister(struct rtnl_link_ops *ops);
 
 int rtnl_link_register(struct rtnl_link_ops *ops);
 void rtnl_link_unregister(struct rtnl_link_ops *ops);
diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index 3b33810d92a8..634732fe4c64 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c
@@ -568,27 +568,6 @@ static void __rtnl_kill_links(struct net *net, struct rtnl_link_ops *ops)
 	unregister_netdevice_many(&list_kill);
 }
 
-/**
- * __rtnl_link_unregister - Unregister rtnl_link_ops from rtnetlink.
- * @ops: struct rtnl_link_ops * to unregister
- *
- * The caller must hold the rtnl_mutex and guarantee net_namespace_list
- * integrity (hold pernet_ops_rwsem for writing to close the race
- * with setup_net() and cleanup_net()).
- */
-void __rtnl_link_unregister(struct rtnl_link_ops *ops)
-{
-	struct net *net;
-
-	list_del_rcu(&ops->list);
-	synchronize_srcu(&ops->srcu);
-	cleanup_srcu_struct(&ops->srcu);
-
-	for_each_net(net)
-		__rtnl_kill_links(net, ops);
-}
-EXPORT_SYMBOL_GPL(__rtnl_link_unregister);
-
 /* Return with the rtnl_lock held when there are no network
  * devices unregistering in any network namespace.
  */
@@ -617,10 +596,19 @@ static void rtnl_lock_unregistering_all(void)
  */
 void rtnl_link_unregister(struct rtnl_link_ops *ops)
 {
+	struct net *net;
+
 	/* Close the race with setup_net() and cleanup_net() */
 	down_write(&pernet_ops_rwsem);
 	rtnl_lock_unregistering_all();
-	__rtnl_link_unregister(ops);
+
+	list_del_rcu(&ops->list);
+	synchronize_srcu(&ops->srcu);
+	cleanup_srcu_struct(&ops->srcu);
+
+	for_each_net(net)
+		__rtnl_kill_links(net, ops);
+
 	rtnl_unlock();
 	up_write(&pernet_ops_rwsem);
 }
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 02/10] rtnetlink: Protect link_ops by mutex.

[Kuniyuki Iwashima]

rtnl_link_unregister() holds RTNL and calls synchronize_srcu(),
but rtnl_newlink() will acquire SRCU frist and then RTNL.

Then, we need to unlink ops and call synchronize_srcu() outside
of RTNL to avoid the deadlock.

   rtnl_link_unregister()       rtnl_newlink()
   ----                         ----
   lock(rtnl_mutex);
                                lock(&ops->srcu);
                                lock(rtnl_mutex);
   sync(&ops->srcu);

Let's move as such and add a mutex to protect link_ops.

Now, link_ops is protected by its dedicated mutex and
rtnl_link_register() no longer needs to hold RTNL.

While at it, we move the initialisation of ops->dellink and
ops->srcu out of the mutex scope.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 include/net/rtnetlink.h |  2 +-
 net/core/rtnetlink.c    | 33 ++++++++++++++++++++-------------
 2 files changed, 21 insertions(+), 14 deletions(-)

diff --git a/include/net/rtnetlink.h b/include/net/rtnetlink.h
index 3ebfcc6e56fd..7559020f760c 100644
--- a/include/net/rtnetlink.h
+++ b/include/net/rtnetlink.h
@@ -71,7 +71,7 @@ static inline int rtnl_msg_family(const struct nlmsghdr *nlh)
 /**
  *	struct rtnl_link_ops - rtnetlink link operations
  *
- *	@list: Used internally, protected by RTNL and SRCU
+ *	@list: Used internally, protected by link_ops_mutex and SRCU
  *	@srcu: Used internally
  *	@kind: Identifier
  *	@netns_refund: Physical device, move to init_net on netns exit
diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index 634732fe4c64..fcccb916e468 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c
@@ -466,6 +466,7 @@ void __rtnl_unregister_many(const struct rtnl_msg_handler *handlers, int n)
 }
 EXPORT_SYMBOL_GPL(__rtnl_unregister_many);
 
+static DEFINE_MUTEX(link_ops_mutex);
 static LIST_HEAD(link_ops);
 
 static struct rtnl_link_ops *rtnl_link_ops_get(const char *kind, int *srcu_index)
@@ -508,14 +509,6 @@ int __rtnl_link_register(struct rtnl_link_ops *ops)
 	struct rtnl_link_ops *tmp;
 	int err;
 
-	/* When RTNL is removed, add lock for link_ops. */
-	ASSERT_RTNL();
-
-	list_for_each_entry(tmp, &link_ops, list) {
-		if (!strcmp(ops->kind, tmp->kind))
-			return -EEXIST;
-	}
-
 	/* The check for alloc/setup is here because if ops
 	 * does not have that filled up, it is not possible
 	 * to use the ops for creating device. So do not
@@ -528,9 +521,20 @@ int __rtnl_link_register(struct rtnl_link_ops *ops)
 	if (err)
 		return err;
 
+	mutex_lock(&link_ops_mutex);
+
+	list_for_each_entry(tmp, &link_ops, list) {
+		if (!strcmp(ops->kind, tmp->kind)) {
+			err = -EEXIST;
+			goto unlock;
+		}
+	}
+
 	list_add_tail_rcu(&ops->list, &link_ops);
+unlock:
+	mutex_unlock(&link_ops_mutex);
 
-	return 0;
+	return err;
 }
 EXPORT_SYMBOL_GPL(__rtnl_link_register);
 
@@ -598,14 +602,17 @@ void rtnl_link_unregister(struct rtnl_link_ops *ops)
 {
 	struct net *net;
 
-	/* Close the race with setup_net() and cleanup_net() */
-	down_write(&pernet_ops_rwsem);
-	rtnl_lock_unregistering_all();
-
+	mutex_lock(&link_ops_mutex);
 	list_del_rcu(&ops->list);
+	mutex_unlock(&link_ops_mutex);
+
 	synchronize_srcu(&ops->srcu);
 	cleanup_srcu_struct(&ops->srcu);
 
+	/* Close the race with setup_net() and cleanup_net() */
+	down_write(&pernet_ops_rwsem);
+	rtnl_lock_unregistering_all();
+
 	for_each_net(net)
 		__rtnl_kill_links(net, ops);
 
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 03/10] rtnetlink: Remove __rtnl_link_register()

[Kuniyuki Iwashima]

link_ops is protected by link_ops_mutex and no longer needs RTNL,
so we have no reason to have __rtnl_link_register() separately.

Let's remove it and call rtnl_link_register() from ifb.ko and
dummy.ko.

Note that both modules' init() work on init_net only, so we need
not export pernet_ops_rwsem and can use rtnl_net_lock() there.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
---
 drivers/net/dummy.c      | 17 ++++++-----------
 drivers/net/ifb.c        | 17 ++++++-----------
 include/net/rtnetlink.h  |  2 --
 net/core/net_namespace.c |  1 -
 net/core/rtnetlink.c     | 35 +++++++----------------------------
 5 files changed, 19 insertions(+), 53 deletions(-)

diff --git a/drivers/net/dummy.c b/drivers/net/dummy.c
index 72618b6af44e..005d79975f3b 100644
--- a/drivers/net/dummy.c
+++ b/drivers/net/dummy.c
@@ -166,27 +166,22 @@ static int __init dummy_init_one(void)
 
 static int __init dummy_init_module(void)
 {
-	bool need_unregister = false;
 	int i, err = 0;
 
-	down_write(&pernet_ops_rwsem);
-	rtnl_lock();
-	err = __rtnl_link_register(&dummy_link_ops);
+	err = rtnl_link_register(&dummy_link_ops);
 	if (err < 0)
-		goto out;
+		return err;
+
+	rtnl_net_lock(&init_net);
 
 	for (i = 0; i < numdummies && !err; i++) {
 		err = dummy_init_one();
 		cond_resched();
 	}
-	if (err < 0)
-		need_unregister = true;
 
-out:
-	rtnl_unlock();
-	up_write(&pernet_ops_rwsem);
+	rtnl_net_unlock(&init_net);
 
-	if (need_unregister)
+	if (err < 0)
 		rtnl_link_unregister(&dummy_link_ops);
 
 	return err;
diff --git a/drivers/net/ifb.c b/drivers/net/ifb.c
index a4b9ec4e8f30..67424888ff0a 100644
--- a/drivers/net/ifb.c
+++ b/drivers/net/ifb.c
@@ -424,27 +424,22 @@ static int __init ifb_init_one(int index)
 
 static int __init ifb_init_module(void)
 {
-	bool need_unregister = false;
 	int i, err;
 
-	down_write(&pernet_ops_rwsem);
-	rtnl_lock();
-	err = __rtnl_link_register(&ifb_link_ops);
+	err = rtnl_link_register(&ifb_link_ops);
 	if (err < 0)
-		goto out;
+		return err;
+
+	rtnl_net_lock(&init_net);
 
 	for (i = 0; i < numifbs && !err; i++) {
 		err = ifb_init_one(i);
 		cond_resched();
 	}
-	if (err)
-		need_unregister = true;
 
-out:
-	rtnl_unlock();
-	up_write(&pernet_ops_rwsem);
+	rtnl_net_unlock(&init_net);
 
-	if (need_unregister)
+	if (err)
 		rtnl_link_unregister(&ifb_link_ops);
 
 	return err;
diff --git a/include/net/rtnetlink.h b/include/net/rtnetlink.h
index 7559020f760c..ef7c11f0d74c 100644
--- a/include/net/rtnetlink.h
+++ b/include/net/rtnetlink.h
@@ -164,8 +164,6 @@ struct rtnl_link_ops {
 						   int *prividx, int attr);
 };
 
-int __rtnl_link_register(struct rtnl_link_ops *ops);
-
 int rtnl_link_register(struct rtnl_link_ops *ops);
 void rtnl_link_unregister(struct rtnl_link_ops *ops);
 
diff --git a/net/core/net_namespace.c b/net/core/net_namespace.c
index 809b48c0a528..157021ced442 100644
--- a/net/core/net_namespace.c
+++ b/net/core/net_namespace.c
@@ -56,7 +56,6 @@ static bool init_net_initialized;
  * outside.
  */
 DECLARE_RWSEM(pernet_ops_rwsem);
-EXPORT_SYMBOL_GPL(pernet_ops_rwsem);
 
 #define MIN_PERNET_OPS_ID	\
 	((sizeof(struct net_generic) + sizeof(void *) - 1) / sizeof(void *))
diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index fcccb916e468..61bf710f97b8 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c
@@ -495,20 +495,21 @@ static void rtnl_link_ops_put(struct rtnl_link_ops *ops, int srcu_index)
 }
 
 /**
- * __rtnl_link_register - Register rtnl_link_ops with rtnetlink.
+ * rtnl_link_register - Register rtnl_link_ops with rtnetlink.
  * @ops: struct rtnl_link_ops * to register
  *
- * The caller must hold the rtnl_mutex. This function should be used
- * by drivers that create devices during module initialization. It
- * must be called before registering the devices.
- *
  * Returns 0 on success or a negative error code.
  */
-int __rtnl_link_register(struct rtnl_link_ops *ops)
+int rtnl_link_register(struct rtnl_link_ops *ops)
 {
 	struct rtnl_link_ops *tmp;
 	int err;
 
+	/* Sanity-check max sizes to avoid stack buffer overflow. */
+	if (WARN_ON(ops->maxtype > RTNL_MAX_TYPE ||
+		    ops->slave_maxtype > RTNL_SLAVE_MAX_TYPE))
+		return -EINVAL;
+
 	/* The check for alloc/setup is here because if ops
 	 * does not have that filled up, it is not possible
 	 * to use the ops for creating device. So do not
@@ -536,28 +537,6 @@ int __rtnl_link_register(struct rtnl_link_ops *ops)
 
 	return err;
 }
-EXPORT_SYMBOL_GPL(__rtnl_link_register);
-
-/**
- * rtnl_link_register - Register rtnl_link_ops with rtnetlink.
- * @ops: struct rtnl_link_ops * to register
- *
- * Returns 0 on success or a negative error code.
- */
-int rtnl_link_register(struct rtnl_link_ops *ops)
-{
-	int err;
-
-	/* Sanity-check max sizes to avoid stack buffer overflow. */
-	if (WARN_ON(ops->maxtype > RTNL_MAX_TYPE ||
-		    ops->slave_maxtype > RTNL_SLAVE_MAX_TYPE))
-		return -EINVAL;
-
-	rtnl_lock();
-	err = __rtnl_link_register(ops);
-	rtnl_unlock();
-	return err;
-}
 EXPORT_SYMBOL_GPL(rtnl_link_register);
 
 static void __rtnl_kill_links(struct net *net, struct rtnl_link_ops *ops)
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 04/10] rtnetlink: Introduce struct rtnl_nets and helpers.

[Kuniyuki Iwashima]

rtnl_newlink() needs to hold 3 per-netns RTNL: 2 for a new device
and 1 for its peer.

We will add rtnl_nets_lock() later, which performs the nested locking
based on struct rtnl_nets, which has an array of struct net pointers.

rtnl_nets_add() adds a net pointer to the array and sorts it so that
rtnl_nets_lock() can simply acquire per-netns RTNL from array[0] to [2].

Before calling rtnl_nets_add(), get_net() must be called for the net,
and rtnl_nets_destroy() will call put_net() for each.

Let's apply the helpers to rtnl_newlink().

When CONFIG_DEBUG_NET_SMALL_RTNL is disabled, we do not call
rtnl_net_lock() thus do not care about the array order, so
rtnl_net_cmp_locks() returns -1 so that the loop in rtnl_nets_add()
can be optimised to NOP.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
---
v2:
  * Move struct rtnl_nets to net/core/rtnetlink.c
  * Unexport rtnl_nets_add()
---
 net/core/rtnetlink.c | 70 ++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 67 insertions(+), 3 deletions(-)

diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index 61bf710f97b8..1879a46a9d7e 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c
@@ -258,8 +258,67 @@ bool lockdep_rtnl_net_is_held(struct net *net)
 	return lockdep_rtnl_is_held() && lockdep_is_held(&net->rtnl_mutex);
 }
 EXPORT_SYMBOL(lockdep_rtnl_net_is_held);
+#else
+static int rtnl_net_cmp_locks(const struct net *net_a, const struct net *net_b)
+{
+	/* No need to swap */
+	return -1;
+}
 #endif
 
+struct rtnl_nets {
+	/* ->newlink() needs to freeze 3 netns at most;
+	 * 2 for the new device, 1 for its peer.
+	 */
+	struct net *net[3];
+	unsigned char len;
+};
+
+static void rtnl_nets_init(struct rtnl_nets *rtnl_nets)
+{
+	memset(rtnl_nets, 0, sizeof(*rtnl_nets));
+}
+
+static void rtnl_nets_destroy(struct rtnl_nets *rtnl_nets)
+{
+	int i;
+
+	for (i = 0; i < rtnl_nets->len; i++) {
+		put_net(rtnl_nets->net[i]);
+		rtnl_nets->net[i] = NULL;
+	}
+
+	rtnl_nets->len = 0;
+}
+
+/**
+ * rtnl_nets_add - Add netns to be locked before ->newlink().
+ *
+ * @rtnl_nets: rtnl_nets pointer passed to ->get_peer_net().
+ * @net: netns pointer with an extra refcnt held.
+ *
+ * The extra refcnt is released in rtnl_nets_destroy().
+ */
+static void rtnl_nets_add(struct rtnl_nets *rtnl_nets, struct net *net)
+{
+	int i;
+
+	DEBUG_NET_WARN_ON_ONCE(rtnl_nets->len == ARRAY_SIZE(rtnl_nets->net));
+
+	for (i = 0; i < rtnl_nets->len; i++) {
+		switch (rtnl_net_cmp_locks(rtnl_nets->net[i], net)) {
+		case 0:
+			put_net(net);
+			return;
+		case 1:
+			swap(rtnl_nets->net[i], net);
+		}
+	}
+
+	rtnl_nets->net[i] = net;
+	rtnl_nets->len++;
+}
+
 static struct rtnl_link __rcu *__rcu *rtnl_msg_handlers[RTNL_FAMILY_MAX + 1];
 
 static inline int rtm_msgindex(int msgtype)
@@ -3770,6 +3829,7 @@ static int rtnl_newlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 	struct net *tgt_net, *link_net = NULL;
 	struct rtnl_link_ops *ops = NULL;
 	struct rtnl_newlink_tbs *tbs;
+	struct rtnl_nets rtnl_nets;
 	int ops_srcu_index;
 	int ret;
 
@@ -3813,6 +3873,8 @@ static int rtnl_newlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 #endif
 	}
 
+	rtnl_nets_init(&rtnl_nets);
+
 	if (ops) {
 		if (ops->maxtype > RTNL_MAX_TYPE) {
 			ret = -EINVAL;
@@ -3842,6 +3904,8 @@ static int rtnl_newlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 		goto put_ops;
 	}
 
+	rtnl_nets_add(&rtnl_nets, tgt_net);
+
 	if (tb[IFLA_LINK_NETNSID]) {
 		int id = nla_get_s32(tb[IFLA_LINK_NETNSID]);
 
@@ -3852,6 +3916,8 @@ static int rtnl_newlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 			goto put_net;
 		}
 
+		rtnl_nets_add(&rtnl_nets, link_net);
+
 		if (!netlink_ns_capable(skb, link_net->user_ns, CAP_NET_ADMIN)) {
 			ret = -EPERM;
 			goto put_net;
@@ -3861,9 +3927,7 @@ static int rtnl_newlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 	ret = __rtnl_newlink(skb, nlh, ops, tgt_net, link_net, tbs, data, extack);
 
 put_net:
-	if (link_net)
-		put_net(link_net);
-	put_net(tgt_net);
+	rtnl_nets_destroy(&rtnl_nets);
 put_ops:
 	if (ops)
 		rtnl_link_ops_put(ops, ops_srcu_index);
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 06/10] veth: Set VETH_INFO_PEER to veth_link_ops.peer_type.

[Kuniyuki Iwashima]

For per-netns RTNL, we need to prefetch the peer device's netns.

Let's set rtnl_link_ops.peer_type and accordingly remove duplicated
validation in ->newlink().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
---
 drivers/net/veth.c | 18 ++++--------------
 1 file changed, 4 insertions(+), 14 deletions(-)

diff --git a/drivers/net/veth.c b/drivers/net/veth.c
index 18148e068aa0..0d6d0d749d44 100644
--- a/drivers/net/veth.c
+++ b/drivers/net/veth.c
@@ -1781,19 +1781,11 @@ static int veth_newlink(struct net *src_net, struct net_device *dev,
 	/*
 	 * create and register peer first
 	 */
-	if (data != NULL && data[VETH_INFO_PEER] != NULL) {
-		struct nlattr *nla_peer;
+	if (data && data[VETH_INFO_PEER]) {
+		struct nlattr *nla_peer = data[VETH_INFO_PEER];
 
-		nla_peer = data[VETH_INFO_PEER];
 		ifmp = nla_data(nla_peer);
-		err = rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack);
-		if (err < 0)
-			return err;
-
-		err = veth_validate(peer_tb, NULL, extack);
-		if (err < 0)
-			return err;
-
+		rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack);
 		tbp = peer_tb;
 	} else {
 		ifmp = NULL;
@@ -1809,9 +1801,6 @@ static int veth_newlink(struct net *src_net, struct net_device *dev,
 	}
 
 	net = rtnl_link_get_net(src_net, tbp);
-	if (IS_ERR(net))
-		return PTR_ERR(net);
-
 	peer = rtnl_create_link(net, ifname, name_assign_type,
 				&veth_link_ops, tbp, extack);
 	if (IS_ERR(peer)) {
@@ -1952,6 +1941,7 @@ static struct rtnl_link_ops veth_link_ops = {
 	.newlink	= veth_newlink,
 	.dellink	= veth_dellink,
 	.policy		= veth_policy,
+	.peer_type	= VETH_INFO_PEER,
 	.maxtype	= VETH_INFO_MAX,
 	.get_link_net	= veth_get_link_net,
 	.get_num_tx_queues	= veth_get_num_queues,
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 07/10] vxcan: Set VXCAN_INFO_PEER to vxcan_link_ops.peer_type.

[Kuniyuki Iwashima]

For per-netns RTNL, we need to prefetch the peer device's netns.

Let's set rtnl_link_ops.peer_type and accordingly remove duplicated
validation in ->newlink().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
---
Note for CAN maintainers, this patch needs to go through net-next
directly as the later patch depends on this.
---
 drivers/net/can/vxcan.c | 12 +++---------
 1 file changed, 3 insertions(+), 9 deletions(-)

diff --git a/drivers/net/can/vxcan.c b/drivers/net/can/vxcan.c
index 9e1b7d41005f..da7c72105fb6 100644
--- a/drivers/net/can/vxcan.c
+++ b/drivers/net/can/vxcan.c
@@ -188,14 +188,10 @@ static int vxcan_newlink(struct net *net, struct net_device *dev,
 
 	/* register peer device */
 	if (data && data[VXCAN_INFO_PEER]) {
-		struct nlattr *nla_peer;
+		struct nlattr *nla_peer = data[VXCAN_INFO_PEER];
 
-		nla_peer = data[VXCAN_INFO_PEER];
 		ifmp = nla_data(nla_peer);
-		err = rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack);
-		if (err < 0)
-			return err;
-
+		rtnl_nla_parse_ifinfomsg(peer_tb, nla_peer, extack);
 		tbp = peer_tb;
 	}
 
@@ -208,9 +204,6 @@ static int vxcan_newlink(struct net *net, struct net_device *dev,
 	}
 
 	peer_net = rtnl_link_get_net(net, tbp);
-	if (IS_ERR(peer_net))
-		return PTR_ERR(peer_net);
-
 	peer = rtnl_create_link(peer_net, ifname, name_assign_type,
 				&vxcan_link_ops, tbp, extack);
 	if (IS_ERR(peer)) {
@@ -302,6 +295,7 @@ static struct rtnl_link_ops vxcan_link_ops = {
 	.newlink	= vxcan_newlink,
 	.dellink	= vxcan_dellink,
 	.policy		= vxcan_policy,
+	.peer_type	= VXCAN_INFO_PEER,
 	.maxtype	= VXCAN_INFO_MAX,
 	.get_link_net	= vxcan_get_link_net,
 };
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 08/10] netkit: Set IFLA_NETKIT_PEER_INFO to netkit_link_ops.peer_type.

[Kuniyuki Iwashima]

For per-netns RTNL, we need to prefetch the peer device's netns.

Let's set rtnl_link_ops.peer_type and accordingly remove duplicated
validation in ->newlink().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Nikolay Aleksandrov <razor@blackwall.org>
---
 drivers/net/netkit.c | 11 ++---------
 1 file changed, 2 insertions(+), 9 deletions(-)

diff --git a/drivers/net/netkit.c b/drivers/net/netkit.c
index cd8360b9bbde..bb07725d1c72 100644
--- a/drivers/net/netkit.c
+++ b/drivers/net/netkit.c
@@ -351,12 +351,7 @@ static int netkit_new_link(struct net *src_net, struct net_device *dev,
 		if (data[IFLA_NETKIT_PEER_INFO]) {
 			attr = data[IFLA_NETKIT_PEER_INFO];
 			ifmp = nla_data(attr);
-			err = rtnl_nla_parse_ifinfomsg(peer_tb, attr, extack);
-			if (err < 0)
-				return err;
-			err = netkit_validate(peer_tb, NULL, extack);
-			if (err < 0)
-				return err;
+			rtnl_nla_parse_ifinfomsg(peer_tb, attr, extack);
 			tbp = peer_tb;
 		}
 		if (data[IFLA_NETKIT_SCRUB])
@@ -391,9 +386,6 @@ static int netkit_new_link(struct net *src_net, struct net_device *dev,
 		return -EOPNOTSUPP;
 
 	net = rtnl_link_get_net(src_net, tbp);
-	if (IS_ERR(net))
-		return PTR_ERR(net);
-
 	peer = rtnl_create_link(net, ifname, ifname_assign_type,
 				&netkit_link_ops, tbp, extack);
 	if (IS_ERR(peer)) {
@@ -978,6 +970,7 @@ static struct rtnl_link_ops netkit_link_ops = {
 	.fill_info	= netkit_fill_info,
 	.policy		= netkit_policy,
 	.validate	= netkit_validate,
+	.peer_type	= IFLA_NETKIT_PEER_INFO,
 	.maxtype	= IFLA_NETKIT_MAX,
 };
 
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 09/10] rtnetlink: Convert RTM_NEWLINK to per-netns RTNL.

[Kuniyuki Iwashima]

Now, we are ready to convert rtnl_newlink() to per-netns RTNL;
rtnl_link_ops is protected by SRCU and netns is prefetched in
rtnl_newlink().

Let's register rtnl_newlink() with RTNL_FLAG_DOIT_PERNET and
push RTNL down as rtnl_nets_lock().

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
---
v2: Remove __rtnl_unlock() dance in rtnl_newlink().
---
 net/core/rtnetlink.c | 27 ++++++++++++++++++++++++---
 1 file changed, 24 insertions(+), 3 deletions(-)

diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index fe65dd6cde8d..cac94fb4be18 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c
@@ -319,6 +319,26 @@ static void rtnl_nets_add(struct rtnl_nets *rtnl_nets, struct net *net)
 	rtnl_nets->len++;
 }
 
+static void rtnl_nets_lock(struct rtnl_nets *rtnl_nets)
+{
+	int i;
+
+	rtnl_lock();
+
+	for (i = 0; i < rtnl_nets->len; i++)
+		__rtnl_net_lock(rtnl_nets->net[i]);
+}
+
+static void rtnl_nets_unlock(struct rtnl_nets *rtnl_nets)
+{
+	int i;
+
+	for (i = 0; i < rtnl_nets->len; i++)
+		__rtnl_net_unlock(rtnl_nets->net[i]);
+
+	rtnl_unlock();
+}
+
 static struct rtnl_link __rcu *__rcu *rtnl_msg_handlers[RTNL_FAMILY_MAX + 1];
 
 static inline int rtm_msgindex(int msgtype)
@@ -3906,9 +3926,7 @@ static int rtnl_newlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 		ops = rtnl_link_ops_get(kind, &ops_srcu_index);
 #ifdef CONFIG_MODULES
 		if (!ops) {
-			__rtnl_unlock();
 			request_module("rtnl-link-%s", kind);
-			rtnl_lock();
 			ops = rtnl_link_ops_get(kind, &ops_srcu_index);
 		}
 #endif
@@ -3971,7 +3989,9 @@ static int rtnl_newlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 		}
 	}
 
+	rtnl_nets_lock(&rtnl_nets);
 	ret = __rtnl_newlink(skb, nlh, ops, tgt_net, link_net, tbs, data, extack);
+	rtnl_nets_unlock(&rtnl_nets);
 
 put_net:
 	rtnl_nets_destroy(&rtnl_nets);
@@ -6975,7 +6995,8 @@ static struct pernet_operations rtnetlink_net_ops = {
 };
 
 static const struct rtnl_msg_handler rtnetlink_rtnl_msg_handlers[] __initconst = {
-	{.msgtype = RTM_NEWLINK, .doit = rtnl_newlink},
+	{.msgtype = RTM_NEWLINK, .doit = rtnl_newlink,
+	 .flags = RTNL_FLAG_DOIT_PERNET},
 	{.msgtype = RTM_DELLINK, .doit = rtnl_dellink},
 	{.msgtype = RTM_GETLINK, .doit = rtnl_getlink,
 	 .dumpit = rtnl_dump_ifinfo, .flags = RTNL_FLAG_DUMP_SPLIT_NLM_DONE},
-- 
2.39.5 (Apple Git-154)

[PATCH v3 net-next 10/10] rtnetlink: Register rtnl_dellink() and rtnl_setlink() with RTNL_FLAG_DOIT_PERNET_WIP.

[Kuniyuki Iwashima]

Currently, rtnl_setlink() and rtnl_dellink() cannot be fully converted
to per-netns RTNL due to a lack of handling peer/lower/upper devices in
different netns.

For example, when we change a device in rtnl_setlink() and need to
propagate that to its upper devices, we want to avoid acquiring all netns
locks, for which we do not know the upper limit.

The same situation happens when we remove a device.

rtnl_dellink() could be transformed to remove a single device in the
requested netns and delegate other devices to per-netns work, and
rtnl_setlink() might be ?

Until we come up with a better idea, let's use a new flag
RTNL_FLAG_DOIT_PERNET_WIP for rtnl_dellink() and rtnl_setlink().

This will unblock converting RTNL users where such devices are not related.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
Reviewed-by: Eric Dumazet <edumazet@google.com>
---
 include/net/rtnetlink.h |  1 +
 net/core/rtnetlink.c    | 19 ++++++++++++++++---
 2 files changed, 17 insertions(+), 3 deletions(-)

diff --git a/include/net/rtnetlink.h b/include/net/rtnetlink.h
index bef76abcff8d..bc0069a8b6ea 100644
--- a/include/net/rtnetlink.h
+++ b/include/net/rtnetlink.h
@@ -13,6 +13,7 @@ typedef int (*rtnl_dumpit_func)(struct sk_buff *, struct netlink_callback *);
 enum rtnl_link_flags {
 	RTNL_FLAG_DOIT_UNLOCKED		= BIT(0),
 #define RTNL_FLAG_DOIT_PERNET		RTNL_FLAG_DOIT_UNLOCKED
+#define RTNL_FLAG_DOIT_PERNET_WIP	RTNL_FLAG_DOIT_UNLOCKED
 	RTNL_FLAG_BULK_DEL_SUPPORTED	= BIT(1),
 	RTNL_FLAG_DUMP_UNLOCKED		= BIT(2),
 	RTNL_FLAG_DUMP_SPLIT_NLM_DONE	= BIT(3),	/* legacy behavior */
diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index cac94fb4be18..5d13b39f884b 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c
@@ -3382,6 +3382,7 @@ static int rtnl_setlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 	struct net *net = sock_net(skb->sk);
 	struct nlattr *tb[IFLA_MAX+1];
 	struct net_device *dev = NULL;
+	struct rtnl_nets rtnl_nets;
 	struct net *tgt_net;
 	int err;
 
@@ -3400,6 +3401,12 @@ static int rtnl_setlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 		goto errout;
 	}
 
+	rtnl_nets_init(&rtnl_nets);
+	rtnl_nets_add(&rtnl_nets, get_net(net));
+	rtnl_nets_add(&rtnl_nets, tgt_net);
+
+	rtnl_nets_lock(&rtnl_nets);
+
 	if (ifm->ifi_index > 0)
 		dev = __dev_get_by_index(net, ifm->ifi_index);
 	else if (tb[IFLA_IFNAME] || tb[IFLA_ALT_IFNAME])
@@ -3412,7 +3419,7 @@ static int rtnl_setlink(struct sk_buff *skb, struct nlmsghdr *nlh,
 	else if (!err)
 		err = -ENODEV;
 
-	put_net(tgt_net);
+	rtnl_nets_unlock(&rtnl_nets);
 errout:
 	return err;
 }
@@ -3497,6 +3504,8 @@ static int rtnl_dellink(struct sk_buff *skb, struct nlmsghdr *nlh,
 			return PTR_ERR(tgt_net);
 	}
 
+	rtnl_net_lock(tgt_net);
+
 	if (ifm->ifi_index > 0)
 		dev = __dev_get_by_index(tgt_net, ifm->ifi_index);
 	else if (tb[IFLA_IFNAME] || tb[IFLA_ALT_IFNAME])
@@ -3511,6 +3520,8 @@ static int rtnl_dellink(struct sk_buff *skb, struct nlmsghdr *nlh,
 	else
 		err = -EINVAL;
 
+	rtnl_net_unlock(tgt_net);
+
 	if (netnsid >= 0)
 		put_net(tgt_net);
 
@@ -6997,10 +7008,12 @@ static struct pernet_operations rtnetlink_net_ops = {
 static const struct rtnl_msg_handler rtnetlink_rtnl_msg_handlers[] __initconst = {
 	{.msgtype = RTM_NEWLINK, .doit = rtnl_newlink,
 	 .flags = RTNL_FLAG_DOIT_PERNET},
-	{.msgtype = RTM_DELLINK, .doit = rtnl_dellink},
+	{.msgtype = RTM_DELLINK, .doit = rtnl_dellink,
+	 .flags = RTNL_FLAG_DOIT_PERNET_WIP},
 	{.msgtype = RTM_GETLINK, .doit = rtnl_getlink,
 	 .dumpit = rtnl_dump_ifinfo, .flags = RTNL_FLAG_DUMP_SPLIT_NLM_DONE},
-	{.msgtype = RTM_SETLINK, .doit = rtnl_setlink},
+	{.msgtype = RTM_SETLINK, .doit = rtnl_setlink,
+	 .flags = RTNL_FLAG_DOIT_PERNET_WIP},
 	{.msgtype = RTM_GETADDR, .dumpit = rtnl_dump_all},
 	{.msgtype = RTM_GETROUTE, .dumpit = rtnl_dump_all},
 	{.msgtype = RTM_GETNETCONF, .dumpit = rtnl_dump_all},
-- 
2.39.5 (Apple Git-154)

Re: [PATCH v3 net-next 00/10] rtnetlink: Convert rtnl_newlink() to per-netns RTNL.

[Jakub Kicinski]

On Wed, 6 Nov 2024 18:28:50 -0800 Kuniyuki Iwashima wrote:
> Patch 5 - 8 are to prefetch the peer device's netns in rtnl_newlink().

Patch 5 did not make it to the list :(
-- 
pw-bot: cr

Re: [PATCH v3 net-next 00/10] rtnetlink: Convert rtnl_newlink() to per-netns RTNL.

[Kuniyuki Iwashima]

From: Jakub Kicinski <kuba@kernel.org>
Date: Thu, 7 Nov 2024 15:53:19 -0800
> On Wed, 6 Nov 2024 18:28:50 -0800 Kuniyuki Iwashima wrote:
> > Patch 5 - 8 are to prefetch the peer device's netns in rtnl_newlink().
> 
> Patch 5 did not make it to the list :(

Oh, I didn't notice that... thanks for the reminder !

It's in my gmail inbox at least, so probably something went
wrong on lore :S

Will resend v3.