Re: [PATCH 1/2] net: dsa: RCU-protect dsa_ptr in struct net_device

[Vladimir Oltean <olteanv@gmail.com>]

Hi Alexander,

On Tue, Sep 10, 2024 at 03:03:15PM +0200, A. Sverdlin wrote:
> From: Alexander Sverdlin <alexander.sverdlin@siemens.com>
> 
> There are multiple races of zeroing dsa_ptr in struct net_device (on
> shutdown/remove) against asynchronous dereferences all over the net
> code. Widespread pattern is as follows:
> 
> CPU0					CPU1
> if (netdev_uses_dsa())
> 					dev->dsa_ptr = NULL;
>         dev->dsa_ptr->...
> 
> One of the possible crashes:
> 
> Unable to handle kernel NULL pointer dereference at virtual address 0000000000000010
> CPU: 0 PID: 12 Comm: ksoftirqd/0 Tainted: G O 6.1.99+ #1
> pc : lan9303_rcv
> lr : lan9303_rcv
> Call trace:
>  lan9303_rcv
>  dsa_switch_rcv
>  __netif_receive_skb_list_core
>  netif_receive_skb_list_internal
>  napi_gro_receive
>  fec_enet_rx_napi
>  __napi_poll
>  net_rx_action
> ...
> 
> RCU-protect dsa_ptr and use rcu_dereference() or rtnl_dereference()
> depending on the calling context.
> 
> Rename netdev_uses_dsa() into __netdev_uses_dsa_currently()
> (assumes ether RCU or RTNL lock held) and netdev_uses_dsa_currently()
> variants which better reflect the uselessness of the function's
> return value, which becomes outdated right after the call.
> 
> Fixes: ee534378f005 ("net: dsa: fix panic when DSA master device unbinds on shutdown")
> Cc: stable@vger.kernel.org
> Signed-off-by: Alexander Sverdlin <alexander.sverdlin@siemens.com>
> ---
>  drivers/net/dsa/mt7530.c                    |   3 +-
>  drivers/net/dsa/ocelot/felix.c              |   3 +-
>  drivers/net/dsa/qca/qca8k-8xxx.c            |   3 +-
>  drivers/net/ethernet/broadcom/bcmsysport.c  |   8 +-
>  drivers/net/ethernet/mediatek/airoha_eth.c  |   2 +-
>  drivers/net/ethernet/mediatek/mtk_eth_soc.c |  22 +++--
>  drivers/net/ethernet/mediatek/mtk_ppe.c     |  15 ++-
>  include/linux/netdevice.h                   |   2 +-
>  include/net/dsa.h                           |  36 +++++--
>  include/net/dsa_stubs.h                     |   6 +-
>  net/bridge/br_input.c                       |   2 +-
>  net/core/dev.c                              |   3 +-
>  net/core/flow_dissector.c                   |  19 ++--
>  net/dsa/conduit.c                           |  66 ++++++++-----
>  net/dsa/dsa.c                               |  19 ++--
>  net/dsa/port.c                              |   3 +-
>  net/dsa/tag.c                               |   3 +-
>  net/dsa/tag.h                               |  19 ++--
>  net/dsa/tag_8021q.c                         |  10 +-
>  net/dsa/tag_brcm.c                          |   2 +-
>  net/dsa/tag_dsa.c                           |   8 +-
>  net/dsa/tag_qca.c                           |  10 +-
>  net/dsa/tag_sja1105.c                       |  22 +++--
>  net/dsa/user.c                              | 104 +++++++++++---------
>  net/ethernet/eth.c                          |   2 +-
>  25 files changed, 240 insertions(+), 152 deletions(-)

Thank you for the patch, and I would like you to not give up on it, even
if we will go for a different bug fix for 'stable'.

It's just that it makes me a bit uneasy to have this as the bug fix.
"Stable" kernels are supposedly named as such because you're not backporting
such a major usage pattern change from the core down to all switch drivers
and N other Ethernet drivers which you can't even test. The diffstat
here says it all, I believe.

This kind of change, in principle, is okay for an early net-next opening.
I have it in my tree, have made a few changes to it, will run more tests
on the HW I have, and will let you know over the course of the following
weeks when it reaches a stage that I'm also comfortable with.