Re: CONFIG_ANDROID

[tlhackque <tlhackque@yahoo.com>]

[-- Attachment #1.1: Type: text/plain, Size: 2293 bytes --]

On 30-Jun-22 07:41, Jason A. Donenfeld wrote:
> On Thu, Jun 30, 2022 at 06:47:38AM -0400, tlhackque wrote:
>> FWIW: Having watched the discussion about CONFIG_ANDROID, it occurs to
>> me that there's an alternative for WireGuard that sidesteps the issue.
>>
>>   From the last patcheset, it seems that the only use in WireGuard is to
>> avoid clearing keys on every wake-up.
> No, it clears keys before sleeping.
>
>> So: Why not timestamp key-clear events, and establish a minimum interval?
> Because we don't know when we're going to wake up again, and the
> objective is to maintain forward secrecy.
>
> Jason

Thanks for the explanation.  One more attempt.

If I understand what's happening:

You're really trying to establish a maximum key lifetime - sleep being a 
proxy for "too long to keep using".  On conventional platforms, that's 
been good enough.  On these Android platforms, it's not.

You're clearing the key before sleeping so that after a presumably 
longish time, you'll negotiate a new one.  But on some platforms, the 
sleeps are so frequent that "longish" is inconveniently short.  And the 
renegotiations are expensive.  On those platforms, you don't clear the 
key to avoid the frequent renegotiations.  This keeps the old key in use 
across the sleeps.

Alternatively, why not make the maximum key lifetime explicit.  E.g. On 
all platforms you could set a renegotiate time when a key is 
established, and if it has expired on wake (or on use) trigger 
renegotiation.  This guarantees a maximum key lifetime, independent of 
the frequency or duration of sleeps.  And you don't need to know when 
you'll wake.

If you also want to make sure that the key isn't in memory longer than 
that time (e.g. to avoid capture on a dump or device loss), you could 
also set a timer (of the sort that wakes the CPU from sleep) that clears 
the key at that time.

There are obvious optimizations if necessary.

The point I'm trying to make is that rather than thinking about the 
annoying platform behavior's effect on the implementation, it's probably 
better to think about what WireGuard is really trying to do and express 
it in the implementation.

I hope this perspective helps.  I'll step out of your way now.



[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]