Re: [PATCH net-next 0/5] tls: implement key updates for TLS1.3

[Jakub Kicinski <kuba@kernel.org>]

On Wed, 18 Jan 2023 11:06:25 +0100 Sabrina Dubroca wrote:
> 2023-01-17, 18:03:51 -0800, Jakub Kicinski wrote:
> > On Tue, 17 Jan 2023 14:45:26 +0100 Sabrina Dubroca wrote:  
> > > This adds support for receiving KeyUpdate messages (RFC 8446, 4.6.3
> > > [1]). A sender transmits a KeyUpdate message and then changes its TX
> > > key. The receiver should react by updating its RX key before
> > > processing the next message.
> > > 
> > > This patchset implements key updates by:
> > >  1. pausing decryption when a KeyUpdate message is received, to avoid
> > >     attempting to use the old key to decrypt a record encrypted with
> > >     the new key
> > >  2. returning -EKEYEXPIRED to syscalls that cannot receive the
> > >     KeyUpdate message, until the rekey has been performed by userspace  
> > 
> > Why? We return to user space after hitting a cmsg, don't we?
> > If the user space wants to keep reading with the old key - 🤷️  
> 
> But they won't be able to read anything. Either we don't pause
> decryption, and the socket is just broken when we look at the next
> record, or we pause, and there's nothing to read until the rekey is
> done. I think that -EKEYEXPIRED is better than breaking the socket
> just because a read snuck in between getting the cmsg and setting the
> new key.

IDK, we don't interpret any other content types/cmsgs, and for well
behaved user space there should be no problem (right?).
I'm weakly against, if nobody agrees with me you can keep as is.

> > >  3. passing the KeyUpdate message to userspace as a control message
> > >  4. allowing updates of the crypto_info via the TLS_TX/TLS_RX
> > >     setsockopts
> > > 
> > > This API has been tested with gnutls to make sure that it allows
> > > userspace libraries to implement key updates [2]. Thanks to Frantisek
> > > Krenzelok <fkrenzel@redhat.com> for providing the implementation in
> > > gnutls and testing the kernel patches.  
> > 
> > Please explain why - the kernel TLS is not faster than user space, 
> > the point of it is primarily to enable offload. And you don't add
> > offload support here.  
> 
> Well, TLS1.3 support was added 4 years ago, and yet the offload still
> doesn't support 1.3 at all.

I'm pretty sure some devices support it. None of the vendors could 
be bothered to plumb in the kernel support, yet, tho.
I don't know of anyone supporting rekeying.

> IIRC support for KeyUpdates is mandatory in TLS1.3, so currently the
> kernel can't claim to support 1.3, independent of offloading.

The problem is that we will not be able to rekey offloaded connections.
For Tx it's a non-trivial problem given the current architecture.
The offload is supposed to be transparent, we can't fail the rekey just
because the TLS gotten offloaded.

> Some folks did tests with and without kTLS using nbdcopy and found a
> small but noticeable performance improvement (around 8-10%).