On 10/09/2024 15:04, Sabrina Dubroca wrote:

2024-09-06, 15:19:08 +0200, Antonio Quartulli wrote:

quoted

On 04/09/2024 17:01, Sabrina Dubroca wrote:

quoted

2024-09-04, 14:07:23 +0200, Antonio Quartulli wrote:

quoted

On 02/09/2024 16:42, Sabrina Dubroca wrote:

quoted

2024-08-27, 14:07:52 +0200, Antonio Quartulli wrote:

I can think of a few possibilities if this causes too many unwanted
drops:

   - use a linked list of keys, set the primary instead of swapping, and
     let delete remove the unused key(s) by ID instead of slot

   - decouple the TX and RX keys, which also means you don't really need
     to swap keys (swap for the TX key becomes "set primary", swap on RX
     can become a noop since you check both slots for the correct keyid)
     -- and here too delete becomes based on key ID

   - if cs->mutex becomes a spinlock, take it in the datapath when
     looking up keys. this will make sure we get a consistent view of
     the keys state.

   - come up with a scheme to let the datapath retry the key lookup if
     it didn't find the key it wanted (maybe something like a seqcount,
     or maybe taking the lock and retrying if the lookup failed)

I don't know if options 1 and 2 are possible based on how openvpn (the
protocol and the userspace application) models keys, but they seem a
bit "cleaner" on the datapath side (no locking, no retry). But they
require a different API.

After chewing over all these ideas I think we can summarize the requirements
as follows:

* PRIMARY and SECONDARY in the API is just an abstraction for "KEY FOR TX"
and "THE OTHER KEY";
* SWAP means "mark for TX the key that currently was not marked";
* we only need a pointer to the key for TX;
* key for RX is picked up by key ID;

Therefore, how about having an array large enough to store all key IDs (max
ID is 7):

Right, I forgot about that. For 8 keys, sure, array sounds ok. MACsec
has only 4 and I also implemented it as an array
(include/net/macsec.h: macsec_{tx,rx}_sc, sa array).

thanks for the pointer!

quoted

* array index is the key ID (fast lookup on RX);
* upon SET_KEY we store the provided keys and erase the rest;

I'm not sure about erasing the rest. If you're installing the
secondary ("next primary") key, you can't just erase the current
primary? ("erase the rest" also means you'd only ever have one key in
the array, which contradicts your last item in this list)

Yeah, my point is wrong.
I fooled myself assuming SET_KEY would install two keys each time, but 
this is not the case, so we can't "erase everything else".

What we can do is:
* if SET_KEY wants to install a key in PRIMARY, then we erase the key 
marked as "primary" and we mark the new one as such
* if SET_KEY wants to install a key in SECONDARY, then we erase the key 
without any mark.

This way we simulate writing into slots.

quoted

* upon SET_KEY we store in a new field the index/ID of the PRIMARY (fast
lookup on TX), namely primary_id;

FTR: this is what "marking as primary" means

quoted

* upon SWAP we just save in primary_id the "other" index/ID;

I'm confused by these two. Both SET_KEY and SWAP modify the primary_id?

Yes.
SET_KEY may want to install a key in the primary slot: in that case the 
new key has to be marked as primary immediately.
This normally happens only upon first SET_KEY after connection, when no 
other key exists.

SWAP will switch primary_id to the other key ID.

makes sense?

quoted

* at any given time we will have only two keys in the array.

This needs to be enforced in the SET_KEY implementation, otherwise
SWAP will have inconsistent effects.

yap. I hope what I wrote above about SET_KEY helps clarifying this part.

quoted

It's pretty much like your option 1 and 2, but using an array
indexed by key ID.

Are you decoupling TX and RX keys (so that they can be installed
independently) in this proposal? I can't really tell, the array could
be key pairs or separate.

I would not decouple TX and RX keys.
Each array item should be the same as what we are now storing into each 
slot (struct ovpn_crypto_key_slot *).
I could use two arrays, instead of an array of slots, but that means 
changing way more logic and I don't see the benefit.

quoted

The concept of slot is a bit lost, but it is not important as long as we can
keep the API and its semantics the same.

Would it be a problem for userspace to keep track of key IDs, and use
that (instead of slots) to communicate with the kernel? Make
setkey/delkey be based on keyid, and replace swap with a set_tx
operation which updates the current keyid for TX. That would seem like
a better API, especially for the per-ID array you're proposing
here. The concept of slots would be almost completely lost (only
"current tx key" is left, which is similar to the "primary slot").
(it becomes almost identical to the way MACsec does things, except
possibly that in MACsec the TX and RX keys are not paired at all, so
you can install/delete a TX key without touching any RX key)

unfortunately changing userspace that way is not currently viable.
There are more components (and documentation) that attach to this slot 
abstraction.

Hence my proposal to keep the API the same, but rework the internal for 
better implementation.

Maybe you can also rework the current code to look a bit like this
array-based proposal, and not give up the notion of slots, if that's
something strongly built into the core of openvpn:

- primary/secondary in ovpn_crypto_state become slots[2]
- add a one-bit "primary" reference, used to look into the slots array
- swap just flips that "primary" bit
- TX uses slots[primary]
- RX keeps inspecting both slots (now slot[0] and slot[1] instead of
   primary/secondary) looking for the correct keyid
- setkey/delkey still operate on primary/secondary, but need to figure
   out whether slot[0] or slot[1] is primary/secondary based on
   ->primary

It's almost identical to the current code (and API), except you don't
need to reassign any pointers to swap keys, so it should avoid the
rekey issue.

This approach sounds very close to what I was aiming for, but simpler 
because we have slots[2] instead of slots[8] (which means that 
primary_id can be just 'one-bit').

The only downside is that upon RX we have to check both keys.
However I don't think this is an issue as we have just 2 keys at most.

I could even optimize the lookup by starting always from the primary 
key, as that's the one being used most of the time by the sender.

Ok, I will go with this approach you summarized here at the end.

Thanks a lot!
Cheers,

-- 
Antonio Quartulli
OpenVPN Inc.