On Mon, Feb 17, 2025 at 4:12 PM Ben Greear <greearb@xxxxxxxxxxxxxxx> wrote: > > On 2/17/25 1:28 PM, Tom Talpey wrote: > > On 2/17/2025 4:18 PM, Steve French wrote: > >> On Mon, Feb 17, 2025 at 3:08 PM Tom Talpey <tom@xxxxxxxxxx> wrote: > >>> > >>> On 2/17/2025 1:27 PM, Steve French wrote: > >>>> Noticed this old bug today when cleaning up emails. > >>>> > >>>> When the user specifies a srcaddr on mount, the DNS resolution of the > >>>> host name should only look for the same type of address (ie IPv4 if > >>>> srcaddr is IPv4, IPv6 if IPv6) right? > >>>> > >>>> Any thoughts on how this was handled in other protocols? > >>> > >>> What is this "srcaddr" witchcraft that thou dost utter? :) > >> > >> The original patch which added it was this, and apparently is needed in some > >> cases where the subnet the request comes from is restricted: > >> > >> commit 3eb9a8893a76cf1cda3b41c3212eb2cfe83eae0e > >> Author: Ben Greear <greearb@xxxxxxxxxxxxxxx> > >> Date: Wed Sep 1 17:06:02 2010 -0700 > >> > >> cifs: Allow binding to local IP address. > >> > >> When using multi-homed machines, it's nice to be able to specify > >> the local IP to use for outbound connections. This patch gives > >> cifs the ability to bind to a particular IP address. > >> > >> Usage: mount -t cifs -o srcaddr=192.168.1.50,user=foo, ... > >> Usage: mount -t cifs -o srcaddr=2002::100:1,user=foo, ... > >> > >> Acked-by: Jeff Layton <jlayton@xxxxxxxxxx> > >> Acked-by: Dr. David Holder <david.holder@xxxxxxxxxxx> > >> Signed-off-by: Ben Greear <greearb@xxxxxxxxxxxxxxx> > > > > I still think this is a hack, and unlikely to work reliably. > > Except for that DNS issue, it works as intended as far as I can tell, and someone that doesn't want > the behaviour can just not use it. Presumably this could be fixed in cifs-utils by checking for the cases when "srcaddr=" is specified, only assembling the list of host ip addresses that match IPv4 vs. IPv6 (ie change to resolve_host() function in cifs-utils/resolve_host.c) > I guess we never run CIFS in mixed ipv4/6 environment with DNS. > > >>> There isn't such an option in mount.nfs that I'm aware of. > >>> And, it isn't documented in mount.cifs either. > >> > >> NFS man page does show "clientaddr=" mount option, > >> and it is necessary apparently in some cases (e.g. > >> https://forum.proxmox.com/threads/nfs-mounts-using-wrong-source-ip-interface.70754/) > > > > The NFSv4.0 clientaddr is totally different, that protocol requires > > the client to inform the server which address to establish a > > callback channel to. This horribly broken protocol architecture > > was corrected in NFSv4.1. > > NFS requires a small pile of patches to have a similar behaviour...they are in > our kernel (github.com greearb) in case anyone wants them, but were never accepted upstream. > > Thanks, > Ben > > > > > Tom. > > > >> > >> > >>> It seems like a hack on top of a hack to match the DNS result > >>> to the type of the specified srcaddr. If the server supports > >>> both IP versions and the DNS record exposes them, won't the > >>> same issue occur on "normal" mounts? > >>> > >>> I would not see this playing nicely with multichannel, btw. > >>> Or RDMA. Probably other scenarios. > >>> > >>> Tom. > >>> > >>> > >>>> > >>>> https://bugzilla.kernel.org/show_bug.cgi?id=218523 > >> > > > > > -- > Ben Greear <greearb@xxxxxxxxxxxxxxx> > Candela Technologies Inc http://www.candelatech.com -- Thanks, Steve
