Re: [PATCH] RDMA: don't ignore client->add() failures

[Leon Romanovsky <leon@kernel.org>]

On Sat, Apr 01, 2023 at 04:30:08PM +0900, Tetsuo Handa wrote:
> On 2023/04/01 1:23, Jason Gunthorpe wrote:
> > On Sat, Apr 01, 2023 at 01:19:47AM +0900, Tetsuo Handa wrote:
> >> I guess that either dev_net(netdev) is not appropriately initialized or
> >> dev_net(netdev) != &init_net is too restrictive to call ib_unregister_device_queued().
> >> Where is dev_net(netdev) initialized?
> > 
> > Bernard? What is this net ns check for? It seems surprising this would
> > be here
> > 
> 
> Commit bdcf26bf9b3a ("rdma/siw: network and RDMA core interface") implemented
> siw_netdev_event() with
> 
> 	if (dev_net(netdev) != &init_net)
> 		return NOTIFY_OK;
> 
> check. But why this check is needed was not explained.
> Maybe ib_devices_shared_netns is relevant?
> 
> Since network devices created upon/after unshare(CLONE_NEWNET) have network
> namespace other than init_net, this check completely disables siw_netdev_event()
> after unshare(CLONE_NEWNET). Thus, removing this check looks reasonable.

I agree with Jason, this check is not supposed to be in siw in first place
and needs to be removed.

Thanks