[tor-dev] Proposal 319: RELAY_FRAGMENT cells

[Nick Mathewson]

On Thu, May 14, 2020 at 3:15 PM David Goulet <dgoulet at torproject.org> wrote:
>
> On 11 May (16:47:24), Nick Mathewson wrote:
 [...]
> > # Onion service concerns.
> >
> > We allocate a new extension for use in the ESTABLISH_INTRO by onion services,
> > to indicate that they can receive a wide INTRODUCE2 cell.  This extension
> > contains:
> >
> >         struct wide_intro2_ok {
> >           u16 max_len;
> >         }
> >
> > We allocate a new extension for use in the `ESTABLISH_RENDEZVOUS`
> > cell, to indicate acceptance of wide `RENDEZVOUS2` cells.  This
> > extension contains:
> >
> >         struct wide_rend2_ok {
> >           u16 max_len;
> >         }
> >
> > (Note that `ESTABLISH_RENDEZVOUS` cells do not currently have a an
> > extension mechanism.  They should be extended to use the same
> > extension format as `ESTABLISH_INTRO` cells, with extensions placed
> > after the rendezvous cookie.)
>
> Why would a client need to announce wide cells in the ESTABLISH phase as
> opposed to using protover "Relay=N" ?

This is not for announcing support of wide cells -- this is for
reporting a setting for how wide fragmented cells should be.

> The maximum length of a fragmented cell is capped to 2^16 (u16) so we don't
> really need the establish process to inform us of the maximum expected length
> but rather use the max_len in the first fragment?

This all comes back to an earlier part of the proposal:

    Not all lengths up to 65535 are valid lengths for a fragmented
    cell.  Any length under 499 bytes SHOULD cause the circuit
    to close, since that could fit into a non-fragmented RELAY cell.
    Parties SHOULD enforce maximum lengths for cell types that
    they understand.

In other words, I'm imagining that there is a maximum length for each
cell type that is much shorter than 65535, even though we're using two
bytes for the length field.

The extension in the establish_intro cell is to tell the intro point
the longest introduce1 cell that it should accept;  this extension in
the establish_rend cell is to tell the rendezvous point the longest
rendezvous1 cell that it should accept.

Another way we could do this would be with a set of network parameters
to describe the maximum length of each fragmented cell.  Do you think
that would be simpler?

(I can't quite remember why I specified it this way in the first place.)

> Furthermore, ESTABLISH_INTRO has extensions (only 1 as of today) so they could
> also be fragments themselves and thus I'm not sure I see the point of having
> two different ways of "expecting" fragments for the ESTABLISH_* cells and the
> INTRO/RENDEZVOUS cells?

The difference thing here is that everybody can tell which protocols
that a relay supports, but there is no automatic way to tell which
protocols an onion service or client supports.  Since
INTRODUCE2/RENDEZVOUS2 cells are handled by these clients, they need
to get opted into by the relays.

(I'm not sure I understood the question completely.)

> > # Compatibility
> >
> > This proposal will require the allocation of a new 'Relay' protocol version,
> > to indicate understanding of the RELAY_FRAGMENTED command.
>
> Here is a thought about a DoS vector. Here goes:
>
> As an upper limit of 65KB total fragment size, it represents ~126 cells in
> total so I could basically send *125* cells and then stop which will put in
> memory a bit more than 64KB and it will stay there until the last fragment is
> received.
>
> And then I do that on 1000 different circuits bringing the total count in
> memory to 64GB. All stuck there, all "waiting" for the last fragment.
>
> Our OOM would kick in killing circuits but it just seems to me a very easy way
> to continously kick the OOM of a _service_ which is pretty bad side channel.

A few responses here:

First, we shouldn't allow 65535-byte fragmented cells.   The actual
maximum length should be something more like 1024 or 4096 bytes.

Second, we should make sure that when we are reassembling cells, we
use the same buf_t buffers that we use for other stuff.  Our buffers
are timestamped, so we can tell which buffer has had data stalling for
the longest, and we should use that to make sure we're killing off the
right circuits preferentially.

Third, fragments should only be allowed at an onion service for
INTRODUCE2, and those should only come one at a time from each
introduction point, so the number that it's reassembling at the time
will be limited by the number of intro circuits it has open.  It'll be
the the intro points that have to be keeping a bunch of cells in
assembly at once, and be ready to kill off circuits that dawdle too
long.

Does this make more sense?  If so I'll try to clarify it in the proposal.
-- 
Nick