Just finished paper and red pencil WGLC review of rpki-manifests-02.
This may seem a little odd, given that I'm a co-author, but it's how I
usually review my own documents for WGLC anyway, and in this case the
responsible WG chair asked me to post my comments to the list rather
than just sending them to my co-authors.

I found one substantive issue: an incorrect (IMHO) requirement to use
one-off EE certs to sign manifests associated with CA certificates.
Allowing one-off EE certificates is fine, mandating them is wrong.
The manifest must be regenerated every time that something in the
collection changes, so if there are a lot of non-batched changes
within one manifest thisUpdate/nextUpdate cycle, requiring one-off EE
certificates requires listing a lot of old EE certificates in the CRL.
Given that the CA certificates do not cycle anywhere near that
quickly, this seems like an unreasonable requirement.

Rest of review is clarification, no major substantive issues unless
people get worked up about SHOULD vs MUST clarifications.

===

Section 2.1: First paragraph is confusing and flat wrong in a few
places.  I think this stems from attempting to talk about a "CA's"
publication point rather than a "CA certificate's" publication point.
Second paragraph has minor issues along the same lines.  Suggested
rewrite of the entire section:

   In the case of a CA certificate's manifest of its associated
   publication repository in the scope of the Resource Certificate PKI
   (RPKI), the manifest contains the current published certificates
   issued by this CA certificate, the most recent CRL associated with
   the CA certificate, and all objects that are signed using a
   "single-use" EE certificate issued by this CA certificate.

   In the case where an EE certificate has a defined publication
   repository, the EE certificate's manifest contains all published
   objects that have been signed by the EE certificate's key pair.

===

Section 2.2.1.3.2.1: Last sentence of first paragraph assumes one-off
EE cert and is wrong anyway.  Rewrite:

   Any revoked EE certificates used to sign previous manifests will be
   removed from the CRL upon expiration.

===

Section 2.2.1.3.2.1.4: Last sentence requires one-off EE certs.
Rewrite:

   In such a case, if one-off manifest EE certificates are being used,
   when the authority issues the new manifest, it MUST also issue a
   new CRL that includes the EE certificate corresponding to the old
   manifest.

===

Section 2.2.1.4: This could probably be relaxed slightly.  The MUST
requirement is there to ease validation of manifests for CA
certificates; it's not really necessary for manifest of EE
certificates, as the validator already has the EE certificate in hand
by the time it gets to the manifest in that case.  The requirement is
harmless in any case, so I'm ok with leaving this as it stands.

===

Section 3.1: This is is where my fundamental disagreement with the
one-off EE certificate requirement shows up.

Steps 1 and 2 should only require new keypair and EE certificate if
none already exists or if the CA has chosen to use one-off EE
certificates.  Rewrite:

   1.  If no keypair exists, or if using one-off manifest EE
       certificates, generate a key pair, then generate an EE
       certificate for this key pair to enable relying parties to
       verify the signature on the manifest.

Then renumber following steps, since I just collapsed two steps into
one (easier to read than spreading a conditional across two steps). 

Step 2 bullet list:

  First bullet requires an id-ad-signedobject SIA extension in the EE
  certificate.  This is harmless, and could conceivably be useful, but
  is not really required given that the EE certificate will be bundled
  into the CMS anyway.  I'm ok with leaving this as it stands, am just
  pointing it out in case it's an oversight from a previous edit.

  Second bullet mentions IP resources but omits ASN resources.  Both
  should be handled via RFC 3779 inheritance, so rewrite:

       *  This EE certificate SHOULD describe its RFC 3779 address and
          ASN resources using the "inherit" attribute rather than
          explicit description of a resource set.

  Third bullet once again pushes for one-off EE certificates.
  Rewrite:

       *  The validity times of the EE certificate SHOULD exactly match
          the thisUpdate and nextUpdate times of the manifest when
          using a one-off EE certificate.  The validity times of the
          EE certificate MUST encompass the interval from thisUpdate
          to nextUpdate in any case.

Step 6 assumes one-off EE certificates.  Rewrite:


   6.  If using one-off EE certificates, the private key associated
       with the EE certificate SHOULD now be destroyed.

You will no doubt have noted that I'm using the term "one-off" instead
of "single-use" here.  This is not an accident.  "Single-use" has
proven to be ambiguous, in that it could mean a certificate that's
used to sign multiple generations of a single object or a certificate
that's used to sign exactly one generation of that same object.   A
better term than either might be "throw-away certificate", but that
might be perceived as pejorative (which it's not, the technique is
clever when used appropriately).

For an example of the other meaning of "single-use" within this same
document, see the last paragraph of section 3.2.

===

Section 3.2, last paragraph.  Other sense of term "single-use", as
noted above.  In this case it appears to mean "an EE certificate that
has no id-ad-signedObjectRepository accessMethod in its SIA
extension" (ie, the first sentence is a tautology).  Rewrite:

   EE certificates that do not contain an id-ad-signedObjectRepository
   accessMethod in their SIA extensions do not have repository
   publication points.  The object signed by such an EE certificate is
   published in the repository publication point of the CA certificate
   that issued the EE certificate, and is listed in the corresponding
   manifest for this CA certificate.

More generally, at this point I consider all occurances of the term
"single use" (or variants such as "single-use") in this document to be
suspect; in each case I think it's clear from context which meaning
was intended, but given that the same term can be understood in either
of two very different ways, I think it's best to avoid it entirely.

===

Section 3.3:

Fourth bullet item is dead wrong.  Manifest scope is all objects
signed by a particular certificate, not all objects signed by a
particular authority.  A persistant name is indeed required, but it
certainly can derived in part from the public key value, and probably
should be to minimize the risk of manifest collision during key
rollover.  I'd just remove the fourth bullet item entirely.

Fifth bullet item: s/the old CA associated/the old CA certificate
associated/ (the distinction matters, particularly in this paragraph).

===

Section 4:  There are several occurance of a rather odd phrase:

   The issuer SHOULD honour these values in the issued certificate or
   MUST reject the Certificate Request.

I don't know what this means.  Either it's a SHOULD or it's a MUST,
can't have it both ways.  If there's something more subtle going on
here, somebody's going to have to explain it to me, preferably by
rephrasing all of the occurances of this incantation, because it's
totally opaque to me.

===

Section 6.2, last sentence: s/undetected deletions/undetected
deletions or replays/.

===

Section 6.3, second paragraph: s/valid manifest is also/valid manifest
covering the signed objects is also/.

===

Section 6.4: Interaction between manifests and CRLs is complex.  It's
discussed briefly in section 6.6, this section and 6.6 should
reference each other.

Section 6.4, last paragraph: this is not not necessarily publisher
error, the relying party's clock might be wrong.  Absent other data,
either explanation is plausible.

===

Section 6.6: I find the discussion of the interaction between manifest
and CRL here a bit disturbing for the case where the manifest is stale
and the CRL is current.  Everything this section says is correct, but
one could easily come away with the belief that it's better to believe
an old stale copy of the CRL than to believe the current CRL, just
because the (stale) manifest matches the (old, stale) CRL.  I don't
think we want to prefer stale CRLs over current ones when we have both
and both are valid, so there's some clarification needed here.

I haven't figured out yet whether this interaction is complicated
enough to merit its own subsection with inbound pointers from both
section 6.4 and section 6.6.

===

That's it.
_______________________________________________
sidr mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/sidr

Reply via email to