mbox series

[net,0/3] net: core: fix a lockdep splat in the dev_addr_list.

Message ID 20200925181246.25090-1-ap420073@gmail.com
Headers show
Series net: core: fix a lockdep splat in the dev_addr_list. | expand

Message

Taehee Yoo Sept. 25, 2020, 6:12 p.m. UTC
This patchset is to avoid lockdep splat.

When a stacked interface graph is changed, netif_addr_lock() is called
recursively and it internally calls spin_lock_nested().
The parameter of spin_lock_nested() is 'dev->lower_level',
this is called subclass.
The problem of 'dev->lower_level' is that while 'dev->lower_level' is
being used as a subclass of spin_lock_nested(), its value can be changed.
So, spin_lock_nested() would be called recursively with the same
subclass value, the lockdep understands a deadlock.
In order to avoid this, a new variable is needed and it is going to be
used as a parameter of spin_lock_nested().
The first and second patch is a preparation patch for the third patch.
In the third patch, the problem will be fixed.

The first patch is to add __netdev_upper_dev_unlink().
An existed netdev_upper_dev_unlink() is renamed to
__netdev_upper_dev_unlink(). and netdev_upper_dev_unlink()
is added as an wrapper of this function.

The second patch is to add the netdev_nested_priv structure.
netdev_walk_all_{ upper | lower }_dev() pass both private functions
and "data" pointer to handle their own things.
At this point, the data pointer type is void *.
In order to make it easier to expand common variables and functions,
this new netdev_nested_priv structure is added.

The third patch is to add a new variable 'nested_level'
into the net_device structure.
This variable will be used as a parameter of spin_lock_nested() of
dev->addr_list_lock.
Due to this variable, it can avoid lockdep splat.

Taehee Yoo (3):
  net: core: add __netdev_upper_dev_unlink()
  net: core: introduce struct netdev_nested_priv for nested interface
    infrastructure
  net: core: add nested_level variable in net_device

 drivers/infiniband/core/cache.c               |  10 +-
 drivers/infiniband/core/cma.c                 |   9 +-
 drivers/infiniband/core/roce_gid_mgmt.c       |   9 +-
 drivers/infiniband/ulp/ipoib/ipoib_main.c     |   9 +-
 drivers/net/bonding/bond_alb.c                |   9 +-
 drivers/net/bonding/bond_main.c               |  10 +-
 drivers/net/ethernet/intel/ixgbe/ixgbe_main.c |  37 ++--
 .../net/ethernet/mellanox/mlxsw/spectrum.c    |  24 +--
 .../ethernet/mellanox/mlxsw/spectrum_router.c |  11 +-
 .../mellanox/mlxsw/spectrum_switchdev.c       |  10 +-
 drivers/net/ethernet/rocker/rocker_main.c     |   9 +-
 drivers/net/wireless/quantenna/qtnfmac/core.c |  10 +-
 include/linux/netdevice.h                     |  68 ++++++--
 net/bridge/br_arp_nd_proxy.c                  |  26 ++-
 net/bridge/br_vlan.c                          |  20 ++-
 net/core/dev.c                                | 164 +++++++++++++-----
 net/core/dev_addr_lists.c                     |  12 +-
 17 files changed, 318 insertions(+), 129 deletions(-)

Comments

David Miller Sept. 28, 2020, 10:14 p.m. UTC | #1
From: Taehee Yoo <ap420073@gmail.com>
Date: Fri, 25 Sep 2020 18:12:46 +0000

> This patchset is to avoid lockdep splat.
> 
> When a stacked interface graph is changed, netif_addr_lock() is called
> recursively and it internally calls spin_lock_nested().
> The parameter of spin_lock_nested() is 'dev->lower_level',
> this is called subclass.
> The problem of 'dev->lower_level' is that while 'dev->lower_level' is
> being used as a subclass of spin_lock_nested(), its value can be changed.
> So, spin_lock_nested() would be called recursively with the same
> subclass value, the lockdep understands a deadlock.
> In order to avoid this, a new variable is needed and it is going to be
> used as a parameter of spin_lock_nested().
> The first and second patch is a preparation patch for the third patch.
> In the third patch, the problem will be fixed.
> 
> The first patch is to add __netdev_upper_dev_unlink().
> An existed netdev_upper_dev_unlink() is renamed to
> __netdev_upper_dev_unlink(). and netdev_upper_dev_unlink()
> is added as an wrapper of this function.
> 
> The second patch is to add the netdev_nested_priv structure.
> netdev_walk_all_{ upper | lower }_dev() pass both private functions
> and "data" pointer to handle their own things.
> At this point, the data pointer type is void *.
> In order to make it easier to expand common variables and functions,
> this new netdev_nested_priv structure is added.
> 
> The third patch is to add a new variable 'nested_level'
> into the net_device structure.
> This variable will be used as a parameter of spin_lock_nested() of
> dev->addr_list_lock.
> Due to this variable, it can avoid lockdep splat.

Series applied, thank you.