Re: [PATCH] RDMA/core: Fix stale RoCE GIDs during netdev events at registration

[Jason Gunthorpe <jgg@ziepe.ca>]

On Mon, Feb 02, 2026 at 11:20:22PM +0900, Tetsuo Handa wrote:

> > - Event handlers correctly update stale GIDs even when racing with rescan
> 
> I couldn't confirm that this is always true. What happens if rdma_roce_rescan_device()
> is preempted between make_default_gid() and __ib_cache_gid_add(), and NETDEV_CHANGEADDR
> event runs meanwhile? It sounds to me that stale gid is possible because gid value is
> calculated before holding the GID table mutex...
> 
> rdma_roce_rescan_device() {
>   ib_enum_roce_netdev() {
>     enum_all_gids_of_dev_cb() {
>       add_default_gids() {
>         ib_cache_gid_set_default_gid() {
>           make_default_gid(ndev, &gid); // GIDs populated with OLD MAC
>                                                                 ip link set eth4 addr NEW_MAC
>                                                                 NETDEV_CHANGEADDR queued

I thought this was impossible because enum_all_gids_of_dev_cb() holds
the rtnl_lock()?

Jason