Hi folks,
I'm at the Kernel Recipes conference now and got a chance to talk with
DaveM a bit about WireGuard upstreaming. His viewpoint has recently
solidified: in order to go upstream, WireGuard must port to the
existing crypto API, and handle the Zinc project separately. As DaveM
is the upstream network tree maintainer, his opinion is quite
instructive.
I've long resisted the idea of porting to the existing crypto API,
because I think there are serious problems with it, in terms of
primitives, API, performance, and overall safety. I didn't want to
ship WireGuard in a form that I thought was sub-optimal from a
security perspective, since WireGuard is a security-focused project.
But it seems like with or without us, WireGuard will get ported to the
existing crypto API. So it's probably better that we just fully
embrace it, and afterwards work evolutionarily to get Zinc into Linux
piecemeal. I've ported WireGuard already several times as a PoC to the
API and have a decent idea of the ways it can go wrong and generally
how to do it in the least-bad way.
I realize this kind of compromise might come as a disappointment for
some folks. But it's probably better that as a project we remain
intimately involved with our Linux kernel users and the security of
the implementation, rather than slinking away in protest because we
couldn't get it all in at once. So we'll work with upstream, port to
the crypto API, and get the process moving again. We'll pick up the
Zinc work after that's done.
I also understand there might be interested folks out there who enjoy
working with the crypto API quite a bit and would be happy to work on
the WireGuard port. Please do get in touch if you'd like to
collaborate.
Jason
From: "Jason A. Donenfeld" <[email protected]>
Date: Wed, 25 Sep 2019 10:29:45 +0200
> His viewpoint has recently solidified: in order to go upstream,
> WireGuard must port to the existing crypto API, and handle the Zinc
> project separately.
I didn't say "must" anything, I suggested this as a more smoothe
and efficient way forward.
I'm also a bit disappointed that you felt the need to so quickly
make such an explosive posting to the mailing list when we've
just spoken about this amongst ourselves only 20 minutes ago.
Please proceed in a more smoothe and considerate manner for all
parties involved.
Thank you.
From: Bruno Wolff III <[email protected]>
Date: Wed, 25 Sep 2019 04:17:00 -0500
> Are there going to be two branches, one for using the current API and
> one using Zinc?
This is inapproprate to even discuss at this point.
Let's see what the crypto based stuff looks like, evaluate it,
and then decide how to proceed forward.
Thank you.
"Jason A. Donenfeld" <[email protected]> writes:
> Hi folks,
>
> I'm at the Kernel Recipes conference now and got a chance to talk with
> DaveM a bit about WireGuard upstreaming. His viewpoint has recently
> solidified: in order to go upstream, WireGuard must port to the
> existing crypto API, and handle the Zinc project separately. As DaveM
> is the upstream network tree maintainer, his opinion is quite
> instructive.
>
> I've long resisted the idea of porting to the existing crypto API,
> because I think there are serious problems with it, in terms of
> primitives, API, performance, and overall safety. I didn't want to
> ship WireGuard in a form that I thought was sub-optimal from a
> security perspective, since WireGuard is a security-focused project.
>
> But it seems like with or without us, WireGuard will get ported to the
> existing crypto API. So it's probably better that we just fully
> embrace it, and afterwards work evolutionarily to get Zinc into Linux
> piecemeal. I've ported WireGuard already several times as a PoC to the
> API and have a decent idea of the ways it can go wrong and generally
> how to do it in the least-bad way.
>
> I realize this kind of compromise might come as a disappointment for
> some folks. But it's probably better that as a project we remain
> intimately involved with our Linux kernel users and the security of
> the implementation, rather than slinking away in protest because we
> couldn't get it all in at once. So we'll work with upstream, port to
> the crypto API, and get the process moving again. We'll pick up the
> Zinc work after that's done.
On the contrary, kudos on taking the pragmatic route! Much as I have
enjoyed watching your efforts on Zinc, I always thought it was a shame
it had to hold back the upstreaming of WireGuard. So as far as I'm
concerned, doing that separately sounds like the right approach at this
point, and I'll look forward to seeing the patches land :)
-Toke
Are there going to be two branches, one for using the current API and one
using Zinc?
Hi Dave,
On Wed, Sep 25, 2019 at 12:03 PM David Miller <[email protected]> wrote:
> I didn't say "must" anything, I suggested this as a more smoothe
> and efficient way forward.
s/must/should/g? However it's characterized, I think your jugements
and opinions are generally sound, and I intend to put them into
action.
> I'm also a bit disappointed that you felt the need to so quickly
> make such an explosive posting to the mailing list when we've
Explosive? That's certainly not the intent here. The project is
changing direction in a big way. Collaborating with others on the
crypto API will be an important part of that. Announcing the change in
direction, those intentions, a rationale on why it will be okay, and
inviting collaboration is a responsible thing to do at the earliest
opportunity. Better to announce intent early rather than surprise
people or deter potential collaborators by keeping plans secret.
Jason
Hi,
On Wed, Sep 25, 2019 at 10:29:45AM +0200, Jason A. Donenfeld wrote:
> I've long resisted the idea of porting to the existing crypto API,
> because I think there are serious problems with it, in terms of
> primitives, API, performance, and overall safety. I didn't want to
> ship WireGuard in a form that I thought was sub-optimal from a
> security perspective, since WireGuard is a security-focused project.
>
> But it seems like with or without us, WireGuard will get ported to the
> existing crypto API. So it's probably better that we just fully
> embrace it, and afterwards work evolutionarily to get Zinc into Linux
> piecemeal. I've ported WireGuard already several times as a PoC to the
> API and have a decent idea of the ways it can go wrong and generally
> how to do it in the least-bad way.
>
> I realize this kind of compromise might come as a disappointment for
> some folks. But it's probably better that as a project we remain
> intimately involved with our Linux kernel users and the security of
> the implementation, rather than slinking away in protest because we
> couldn't get it all in at once. So we'll work with upstream, port to
> the crypto API, and get the process moving again. We'll pick up the
> Zinc work after that's done.
>
> I also understand there might be interested folks out there who enjoy
> working with the crypto API quite a bit and would be happy to work on
> the WireGuard port. Please do get in touch if you'd like to
> collaborate.
I have some WIP code to port WG to the crypto API, more to get an idea how hard
it would be, though I read you've ported it to the api already. My other
project (btrfs) is going to use blake2 in kernel and for that I'm about to
submit the code, that's where it's also of interest for wg.
My work is at 'github.com/kdave/WireGuard branch lkca-1'. I tried to find a way
how to minimize the impact on current wg code but make it possible to
iteratively extend it to the crypto API.
So, there's some config-time ifdefery to select which crypto functions are
using kernel or zinc api. See wg.git/src/crypto/Kbuild.include at the top,
plus some source ifdefs. I made an example of blake2s port, but only compile
tested.
There are several problems in general that need to be solved on the kernel side
first, before wireguard can work inside the kernel code base:
* missing crypto functions in kernel
* blake2
* curve25519 (missing completely)
* missing generic crypto API callback to use blake_init_key, it's possible to
use only the no-key variant (I have a patch for that, it's really easy but
it's change in API so ...)
The known problem is the cumbersome way to use the crypto functions, eg. for
chacha/poly, I understand the pain and perhaps the reasons to start a fresh
crypto library. I'm afraid the first implementation with current state of
crypto API will be slow, until the API is extended to provide simple ways to
transform buffers without scatterlists, request allocations, locking tfm
context and whatnot.
Feel free to reuse anything from the code if you think it's going the right
direction. I'm not sure if I'll have time to continue with the port but at
least you can consider blake2 on the way upstream.
d.