mpls: Drop RTNL for RTM_NEWROUTE, RTM_DELROUTE, and RTM_GETROUTE.

RTM_NEWROUTE looks up dev under RCU (ip_route_output(),
ipv6_stub->ipv6_dst_lookup_flow(), netdev_get_by_index()),
and each neighbou

mpls: Drop RTNL for RTM_NEWROUTE, RTM_DELROUTE, and RTM_GETROUTE.

RTM_NEWROUTE looks up dev under RCU (ip_route_output(),
ipv6_stub->ipv6_dst_lookup_flow(), netdev_get_by_index()),
and each neighbour holds the refcnt of its dev.

Also, net->mpls.platform_label is protected by a dedicated
per-netns mutex.

Now, no MPLS code depends on RTNL.

Let's drop RTNL for RTM_NEWROUTE, RTM_DELROUTE, and RTM_GETROUTE.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-14-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Protect net->mpls.platform_label with a per-netns mutex.

MPLS (re)uses RTNL to protect net->mpls.platform_label,
but the lock does not need to be RTNL at all.

Let's protect net->mpls.platform

mpls: Protect net->mpls.platform_label with a per-netns mutex.

MPLS (re)uses RTNL to protect net->mpls.platform_label,
but the lock does not need to be RTNL at all.

Let's protect net->mpls.platform_label with a dedicated
per-netns mutex.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-13-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Convert RTM_GETNETCONF to RCU.

mpls_netconf_get_devconf() calls __dev_get_by_index(),
and this only depends on RTNL.

Let's convert mpls_netconf_get_devconf() to RCU and use
dev_get_by_index_r

mpls: Convert RTM_GETNETCONF to RCU.

mpls_netconf_get_devconf() calls __dev_get_by_index(),
and this only depends on RTNL.

Let's convert mpls_netconf_get_devconf() to RCU and use
dev_get_by_index_rcu().

Note that nlmsg_new() is moved ahead to use GFP_KERNEL.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-12-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Convert mpls_dump_routes() to RCU.

mpls_dump_routes() sets fib_dump_filter.rtnl_held to true and
calls __dev_get_by_index() in mpls_valid_fib_dump_req().

This is the only RTNL dependant in mp

mpls: Convert mpls_dump_routes() to RCU.

mpls_dump_routes() sets fib_dump_filter.rtnl_held to true and
calls __dev_get_by_index() in mpls_valid_fib_dump_req().

This is the only RTNL dependant in mpls_dump_routes().

Also, synchronize_rcu() in resize_platform_label_table()
guarantees that net->mpls.platform_label is alive under RCU.

Let's convert mpls_dump_routes() to RCU and use dev_get_by_index_rcu().

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-11-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Use mpls_route_input() where appropriate.

In many places, we uses rtnl_dereference() twice for
net->mpls.platform_label and net->mpls.platform_label[index].

Let's replace the code with mpls_r

mpls: Use mpls_route_input() where appropriate.

In many places, we uses rtnl_dereference() twice for
net->mpls.platform_label and net->mpls.platform_label[index].

Let's replace the code with mpls_route_input().

We do not use mpls_route_input() in mpls_dump_routes() since
we will rely on RCU there.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-10-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Add mpls_route_input().

mpls_route_input_rcu() is called from mpls_forward() and
mpls_getroute().

The former is under RCU, and the latter is under RTNL, so
mpls_route_input_rcu() uses rcu_der

mpls: Add mpls_route_input().

mpls_route_input_rcu() is called from mpls_forward() and
mpls_getroute().

The former is under RCU, and the latter is under RTNL, so
mpls_route_input_rcu() uses rcu_dereference_rtnl().

Let's use rcu_dereference() in mpls_route_input_rcu() and
add an RTNL variant for mpls_getroute().

Later, we will remove rtnl_dereference() there.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-9-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Pass net to mpls_dev_get().

We will replace RTNL with a per-netns mutex to protect dev->mpls_ptr.

Then, we will use rcu_dereference_protected() with the lockdep_is_held()
annotation, which re

mpls: Pass net to mpls_dev_get().

We will replace RTNL with a per-netns mutex to protect dev->mpls_ptr.

Then, we will use rcu_dereference_protected() with the lockdep_is_held()
annotation, which requires net to access the per-netns mutex.

However, dev_net(dev) is not safe without RTNL.

Let's pass net to mpls_dev_get().

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-8-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Add mpls_dev_rcu().

mpls_dev_get() uses rcu_dereference_rtnl() to fetch dev->mpls_ptr.

We will replace RTNL with a dedicated mutex to protect the field.

Then, we will use rcu_dereference_pro

mpls: Add mpls_dev_rcu().

mpls_dev_get() uses rcu_dereference_rtnl() to fetch dev->mpls_ptr.

We will replace RTNL with a dedicated mutex to protect the field.

Then, we will use rcu_dereference_protected() for clarity.

Let's add mpls_dev_rcu() for the RCU reader.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-7-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Use in6_dev_rcu() and dev_net_rcu() in mpls_forward() and mpls_xmit().

mpls_forward() and mpls_xmit() are called under RCU.

Let's use in6_dev_rcu() and dev_net_rcu() there to annotate
as such

mpls: Use in6_dev_rcu() and dev_net_rcu() in mpls_forward() and mpls_xmit().

mpls_forward() and mpls_xmit() are called under RCU.

Let's use in6_dev_rcu() and dev_net_rcu() there to annotate
as such.

Now we pass net to mpls_stats_inc_outucastpkts() not to read
dev_net_rcu() twice.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-6-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Unify return paths in mpls_dev_notify().

We will protect net->mpls.platform_label by a dedicated mutex.

Then, we need to wrap functions called from mpls_dev_notify()
with the mutex.

As a pre

mpls: Unify return paths in mpls_dev_notify().

We will protect net->mpls.platform_label by a dedicated mutex.

Then, we need to wrap functions called from mpls_dev_notify()
with the mutex.

As a prep, let's unify the return paths.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-4-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Hold dev refcnt for mpls_nh.

MPLS uses RTNL

1) to guarantee the lifetime of struct mpls_nh.nh_dev
2) to protect net->mpls.platform_label

, but neither actually requires RTNL.

If we do n

mpls: Hold dev refcnt for mpls_nh.

MPLS uses RTNL

1) to guarantee the lifetime of struct mpls_nh.nh_dev
2) to protect net->mpls.platform_label

, but neither actually requires RTNL.

If we do not call dev_put() in find_outdev() and call it
just before freeing struct mpls_route, we can drop RTNL for 1).

Let's hold the refcnt of mpls_nh.nh_dev and track it with
netdevice_tracker.

Two notable changes:

If mpls_nh_build_multi() fails to set up a neighbour, we need
to call netdev_put() for successfully created neighbours in
mpls_rt_free_rcu(), so the number of neighbours (rt->rt_nhn)
is now updated in each iteration.

When a dev is unregistered, mpls_ifdown() clones mpls_route
and replaces it with the clone, so the clone requires extra
netdev_hold().

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-3-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...

mpls: Return early in mpls_label_ok().

When mpls_label_ok() returns false, it does not need to update *index.

Let's remove is_ok and return early.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.c

mpls: Return early in mpls_label_ok().

When mpls_label_ok() returns false, it does not need to update *index.

Let's remove is_ok and return early.

Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Link: https://patch.msgid.link/20251029173344.2934622-2-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

show more ...