[PATCH v5 net-next 0/8] vxlan: convert configuration to RCU and enable lockless dumps

[Eric Dumazet <edumazet@google.com>]

This patch series converts struct vxlan_config to RCU protection, eliminates
redundant destination fields, and enables lockless RTNL-free link dumping
in vxlan_fill_info().

Changes in v5:
- Widen vxlan_multicast_leave() and vxlan_multicast_join() conditions in
  vxlan_changelink() to also run when VXLAN_F_VNIFILTER is set, so per-VNI
  multicast memberships are migrated across lower-device changes even when
  the device-default remote_ip is 0.0.0.0 (patches 1, 6, 7).
- Unwind already-updated per-VNI entries and the device-wide default FDB
  entry if vxlan_vnilist_update_group() fails partway through in
  vxlan_changelink() (patches 1, 7).
- Bump the vnifilter dump generation counter inside vxlan_vni_update_group()
  as soon as vninode->remote_ip is overwritten, so concurrent dumps still
  observe NLM_F_DUMP_INTR if a subsequent IGMP leave/join fails (patch 3).
- In vxlan_config_validate(), reject clearing IFLA_VXLAN_LINK when
  VXLAN_F_VNIFILTER has per-VNI multicast groups configured (patch 7).
- In vxlan_config_apply(), reset dev->needed_tailroom when detaching a
  lower device and use READ_ONCE()/WRITE_ONCE() for dev->mtu (patch 7).
- Link to v4: https://lore.kernel.org/netdev/20260915175501.391567-1-edumazet@google.com/

Changes in v4:
- Split out the default FDB update fix when the lower device changes into
  patch 1 with a Fixes: tag, and updated vxlan_vnilist_update_group() to
  refresh all VNIs (including those with explicit per-VNI multicast groups)
  so their FDB entries do not retain a stale remote_ifindex. Kept in this
  series rather than routing through net to avoid a cross-tree conflict
  with the default_dst removal in patch 7.
- Added patch 3 to signal interrupted RTM_GETTUNNEL dumps to user space via
  a per-netns atomic generation counter and nl_dump_check_consistent(), per
  Paolo's suggestion.
- Added an explicit VXLAN_DEV_F_MDB check in mlxsw_sp_nve_vxlan_can_offload()
  so moving VXLAN_F_MDB out of cfg->flags does not silently relax mlxsw's
  deny-by-default flag validation (patch 5).
- In vxlan_config_apply(), gated the MTU clamp on lowerdev_changed so that
  unrelated changelink operations do not shrink dev->mtu (patch 7).
- Documented in patch 7's changelog that cfg->remote_ifindex is now committed
  unconditionally and that IFLA_VXLAN_LINK=0 tears down the upper/lower
  adjacency.
- Link to v3: https://lore.kernel.org/netdev/20260911033945.164207-1-edumazet@google.com/

Changes in v3:
- Folded configuration conversion patches into a single patch to maintain
  clean bisectability and avoid intermediate mutable state.
- In vxlan_changelink(), pass lowerdev to vxlan_config_apply() to prevent
  truncating dev->needed_headroom and dev->needed_tailroom when lowerdev
  does not change.
- Cleaned up redundant NULL checks and unreachable error paths.

Eric Dumazet (8):
  vxlan: update default fdb entries when the lower device changes
  vxlan: vnifilter: use list_for_each_entry_rcu() in
    vxlan_vnifilter_dump_dev()
  vxlan: vnifilter: signal interrupted RTM_GETTUNNEL dumps
  vxlan: pass vxlan_config pointer to helper functions
  vxlan: move VXLAN_F_MDB to struct vxlan_dev flags
  vxlan: convert configuration to RCU protection
  vxlan: remove default_dst and use vxlan_config and lowerdev
  vxlan: no longer rely on RTNL in vxlan_fill_info()

 .../mellanox/mlx5/core/en/tc_tun_vxlan.c      |  11 +-
 .../mellanox/mlxsw/spectrum_nve_vxlan.c       |  19 +-
 .../mellanox/mlxsw/spectrum_switchdev.c       |  57 +-
 drivers/net/vxlan/vxlan_core.c                | 754 +++++++++++-------
 drivers/net/vxlan/vxlan_mdb.c                 |  45 +-
 drivers/net/vxlan/vxlan_multicast.c           |  76 +-
 drivers/net/vxlan/vxlan_private.h             |  34 +-
 drivers/net/vxlan/vxlan_vnifilter.c           | 191 +++--
 include/net/vxlan.h                           |  12 +-
 9 files changed, 769 insertions(+), 430 deletions(-)

-- 
2.55.0.1082.g2b9226bbc0-goog