|
Re: LC comments on draft-laurie-pki-sunlight-05
2013-01-25 05:17:40
Apologies for responding to recent comments in random order: I'm
travelling and have accumulated something of a backlog.
On 22 January 2013 03:11, =JeffH
<Jeff(_dot_)Hodges(_at_)kingsmountain(_dot_)com> wrote:
apologies for latency, many meetings and a conference in the last couple of
weeks.
BenL replied:
On 1 January 2013 21:50, =JeffH
<Jeff(_dot_)Hodges(_at_)kingsmountain(_dot_)com> wrote:
[ in the below discussion:
"the spec", "this spec" refers to draft-laurie-pki-sunlight-05.
"TLS-CT client" refers to a TLS client capable of processing CT information
that is included in the TLS handshake in any of the specified manners.
"ok" means in general: "ok, will check this in next rev of the spec..".
]
<snip>
comments on draft-laurie-pki-sunlight-05
substantive comments (in somewhat arbitrary order)
--------------------------------------------------
[ I demoted the comments wrt "JSON object" terminology and put them down at
the end of this msg ]
Also, the syntax for GETS isn't fully specified. Are the URL parameters
to
be encoded as (order independent) key-value pairs, or just as
order-dependent values? Which separator character is to be used between
parameters? RFC3986 should be cited.
RFC 3986 says nothing about parameter format, though
correct, it doesn't, and I wasn't trying to imply that it did, sorry. I was
just trying to say that RFC3986 should be cited in this spec because this
spec normatively employs URLs, but perhaps referencing RFC2616 HTTP is
better because it defines "http_URL".
ok.
- is there a
standard reference for that? I've refereced HTML 4.01, but perhaps
there's a better one?
hm, AFAICT, there is not a standard for URI query component formating and
thus parameter encoding, so this spec will have to explicitly specify
something. Section 3.4 of RFC3986 gives allowed chars for the query
component, but that's about it.
Have you mocked up code that parses the log client messages? If so, what
query component syntax does it handle?
I have specified the "standard" format via HTML 4.01.
3. There appear to be three defined methods for TLS servers to provide
TLS
clients with CT data, in S3.2. For this experiment, which approach is
mandatory to implement for servers and clients? Or, is it the case that
participating TLS clients (ie web browsers etc) implement all three
methods,
and TLS servers can choose any of them?
The latter.
That should be made very clear. Is the reason for doing so to obtain
operational experience wrt the three defined methods such that they perhaps
can be narrowed down in the future, or is the expectation that TLS-CT
clients will need to support all three methods in perpetuity?
I think I made that clear.
The reason three methods exist are as follows (I don't intend to get
into this in the RFC, but for your edification):
1. TLS extension is the right way to do it, but requires a server s/w
change - this adds many years to full deployment.
2. So, alternatives must be provided. One is to put the stuff in the
certificate, but...
3. ... some CAs have said they'd rather not gate issuance on the log,
so alternatively, wedge the stuff in an OCSP response (which must be
stapled - servers exist that support this option already).
5. The recursive equations in S2.1 describe how to calculate a Merkle
Tree
Hash (MTH) (aka "root hash"), and thus as a side effect generate a Merkle
Tree, for a given set of input data. However, there doesn't seem to be a
defined algorithm (or even hints, really) for adding further inputs to an
existing tree. Even though this may be reasonably left as an exercise for
implementers, it should probably be discussed to some degree in the spec.
E.g., note that leaf hashes are "frozen" and various interior tree node
hashes become "frozen" as the tree grows. Is it not sub-optimal to employ
the obvious default brute-force mechanism of rebuilding a tree entirely
from
scratch when new inputs are available? Would not a recursive algorithm
for
adding new inputs to an existing tree be straightforward to provide?
I dunno about straightforward.
yeah, agreed.
I'll think about it.
ok. At least providing some hints would be useful it seems.
We do provide working code :-)
6. Signed tree heads (STHs) are denoted in terms of "tree size" (number
of
entries), but SCTs are denoted in terms of a timestamp. Should there be
a
log client message supporting the return of the nearest STH (and thus
tree
size) to a given timestamp?
I'm not sure why? Any STH (that includes that SCT) will do.
Hm, it was sort of a gut feel that it might be useful, but perhaps not.
S5.2. Auditor says..
A certificate accompanied by an SCT can be verified against any STH
dated after the SCT timestamp + the Maximum Merge Delay by requesting
a Merkle Audit Proof using Section 4.5.
S4.5 get-proof-by-hash stipulates tree_size as an input, but
if a log auditor doesn't already have tree_size, then I suppose it first
calls S4.3 get-sth, which will return a timestamp and a tree_size, which if
generated max merge delay (MMD) after the SCT was gen'd, ought to be
sufficient, yes?
Yes.
I don't see in the spec where/how MMD is published. Does MMD vary per log
service? The latter isn't stipulated in the spec it seems AFAICT ?
We have not really figured out how MMD is specified. I suspect it is
something that will be agreed between browser vendors and logs.
7. S3 paragraph 2 states that "TLS clients MUST reject certificates that
do
not have a valid SCT for the end-entity certificate" (i.e., hard-fail).
Presummably this requirement is only for TLS clients participating in the
CT
experiment and that understand this protocol.
Of course - what other way could it be? In other words, all RFCs can
only say what implementations that conform with them do.
This, or whatever the
requirement actually is, should be further explained.
I guess what I was getting at is that CT-conformant TLS clients should be
differentiated from non-CT-conformant ones. Stipulating a name for
CT-conformant TLS clients would clarify this (seems to me), e.g. "TLS-CT
client" or something similar.
I understand what you're getting at, but not why. CT conformant TLS
clients obey all the MUSTs and non-conformant ones do not. That is
what MUST means.
8. The spec implies, but doesn't clearly describe, especially in S3.1,
that
the hashes are "labels" for tree entries, and that given a leaf hash, the
log implementation should be able to look up and present the LogEntry
data
defined in that section.
We actually only require an entry to be retrievable by hash (in
effect) for the message in 4.7, which we (at least currently) label as
a debugging message - so I am not sure that logs really are required
to be able to do that - certainly the system would work fine if they
couldn't, I believe (other than being unable to provide the debugging
data).
I think what I was getting at was that this is stated at the end of S3.2..
The leaves of the Merkle Tree are the hashes of the corresponding
"MerkleTreeLeaf" structures.
..but it is somewhat confusing because (abstractly) each leaf also contains
(or has a reference to) the LogEntry data (that's defined back in S3.1)
...if I understand this correctly.
I suppose this reflects our own confusion :-)
A LogEntry shows that someone we know can be blamed for the creation
of the certificate - perhaps indirectly. This is desirable because:
a) We want to limit spam, and
b) If someone starts misissuing, we probably want to know who that is...
However, this information is not need to _fix_ the misissuance, it
turns out. It also is not necessarily the information TLS clients may
see (because of cross-signing). So, I think we;re a bit ambivalent
about the precise role of this information.
While in CT, if the input set is an odd number of entries, then the hash
of
the final single leaf is at layer 1, and is calculated as a leaf hash
using
0x00. Thus CT "interior nodes" always have two children, but if the tree
has
an odd number of entries, the rightmost hash at layer 1 ("j" in the
"binary
Merkle tree with 7 leaves" figure) is a leaf node hash rather than an
interior node hash.
O-8. [CrosbyWallach] discusses auditing and gossiping and could be cited
as
a source for further discussion on those topics.
O-9. The notion of "commitments" isn't well defined, and where "add a
commitment to D[k:n]" couldn't "add an interior/intermediate node to
D[k:n}" be used?
Is not the term "commitment" used in [CrosbyWallach] equivalent to the
sha256_root_hash (an STH component) in the spec?
[CrosbyWallach] uses the term "interior node(s)" while the spec uses
"intermediate nodes" (in one place).
CT is not actually derived from [CrosbyWallach], we just mention it as
a useful reference.
yeah, it is useful, it would've been much harder to figure out the spec
without it. Is there an existing paper that the spec is based more closely
on?
No, we started afresh :-)
The differences between the spec and [CrosbyWallach] doesn't seem to be that
large, and so if there aren't more closely matching paper(s) to cite (?), it
may be worth summarizing the differences between the spec and
[CrosbyWallach] e.g. in an appendix.
"commitment" is a term of cryptographic art.
so it is, but its definitions in the literature seem to vary somewhat
depending on context. IIUC, the spec uses "commitment" to refer to the hash
labels of intermediate nodes ('interior nodes' in [CrosbyWallach]) as well
as the root hash, yes?
In S2.1.1, where it says..
We prove that the left subtree entries D[0:k] are consistent
and add a commitment to D[k:n]:
..it's not clear just what "add a commitment" means. add it (an
intermediate-node hash?) to the list of hashes that the recursive algorthm
is constructing?
Overall, I'm finding S2.1.1 pretty difficult to parse and sort out.
E.g., it isn't clear to me how/why/when the apparently boolean 3d parameter
of SUBPROOF is either true or false, and whether there's an implied "if b
then ... else ..." in there somewhere.
Its reasonably hard to define this thing rigorously, and so I'm not
surprised you find it hard to follow - but as I think I've said
before, we did try various different phrasings and did not find an
easier one. Suggestions welcome.
O-10. The recursive algorithms in S2 are dense and take effort to work
through, perhaps adding simplistic example code (in an appendix) which
implements, and/or actually working through the algorithms to arrive at
some
of the audit paths and consistency proofs in S2.1.3, would be helpful.
We have actual working code - would a reference to that be better?
I don't know if it would be better (I've compiled it and am poking through
the code). I sorta think pseudo code in an appendix that would generate the
trees in S2.1.3. would be helpful.
I suspect this is more relevant to a non experimental RFC - right now
I have no evidence that anyone is planning to actually write code
other than us - AFAIK, so far everyone who has played with it has used
our code.
O-14. Detailed comments on S2...
------------------------------
2. Cryptographic components
2.1. Merkle Hash Trees
Logs use a binary Merkle hash tree for efficient auditing. The
hashing algorithm is SHA-256 (note that this is fixed for this
experiment but it is anticipated that each log would be able to
specify a hash algorithm). The input to the Merkle tree hash is a
list of data entries; these entries will be hashed to form the leaves
of the Merkle hash tree. The output is a single 32-byte root hash.
Given an ordered list of n inputs, D[n] = {d(0), d(1), ..., d(n-1)},
the Merkle Tree Hash (MTH) is thus defined as follows:
The hash of an empty list is the hash of an empty string:
MTH({}) = SHA-256().
This MTH({}) construct doesn't appear to be used anywhere else in the
spec
(yes?), and so does it really need mentioning?
If it is not defined, then we cannot represent an empty tree.
yeah, I agree from a mathematical completeness perspective. but I still
found it sort of distracting in that I don't know that it's actually germane
from an implementor's perspective. maybe it is because
The hash of a list with one entry is:
MTH({d(0)}) = SHA-256(0x00 || d(0)).
The immediately above equation is for leaf entries (yes?),
Yes.
where in this
notation n = 1, perhaps it should be stated explicitly:
When n = 1, a leaf entry is denoted, and D[1] = {d(0)}. The leaf hash
(LH) for a leaf entry is calculated as:
MTH(D[1]) = LH(D[1]) = SHA-256( 0x00 || d(0) )
Ugh. LH(D[1]) seems meaningless to me. A leaf hash is always of a "1
entry tree".
yeah, for those who have grokked this stuff. for others it isn't immediately
obvious. also, the above comment was motivated in part due to the missing
formal definition for "Leaf Hash" noted in item (4) above. even if the
LH(D[1]) notation isn't used, some additional prose along the lines
suggested above would be helpful to less initiated readers imv.
For n > 1, let k be the largest power of two smaller than n.
The unqualified "power of two" phrase is arguably ambiguous.
It is?
uh, yeah.
but "let k be a number which is the largest power of two such that..."
isn't.
Suggested rephrase for this where it occurs throughout section 2..
For n > 1, let k be a number which is the largest power of two
such that k = 2^i, 0 <= i < n, and k < n.
If we're going to go down that path, then it should say:
For n > 1, let k be the largest number such that k = 2^i and k < n.
or
For n > 1, let k = 2^i s.t. k < n and 2k >= n.
surely?
well, yes (i prefer the former), but i think it's also important to
explicitly state 0 <= i < n
Why? Its weird! its also not generally true - e.g. n=2 and i=1.
The Merkle Tree Hash of an n-element list D[n] is then defined
recursively as
The above statement applies to the combination of the n = 1 equation
above
and the equation below, and so should perhaps be moved up above the n = 1
equation.
? It says n > 1, so doesn't apply to n = 1?
oh yeah huh.
in any case, I found the prose separation of the "list with one entry" and
the "n > 1" case to be confusing because the former is part of the recursive
definition of MTH(), yes?
OK, so D[0] = {} and D[1] = {d[0]} and we should make that more
explicit as you say which I expect will make this a lot clearer.
MTH(D[n]) = SHA-256(0x01 || MTH(D[0:k]) || MTH(D[k:n])),
where || is concatenation and D[k1:k2] denotes the length (k2 - k1)
list {d(k1), d(k1+1),..., d(k2-1)}.
The above phrase doesn't parse well and is somewhat ambiguous, here it is
extracted for clarity:
"D[k1:k2] denotes the length (k2 - k1) list {d(k1), d(k1+1),...,
d(k2-1)}"
How about rephrasing it along the lines of this:
D[k1:k2] denotes a sublist {d(k1), d(k1+1),..., d(k2-1)}, having
(k2 - k1) elements, of the original input list D[n]. When (k2 - k1)
is 1, a leaf hash is calculated.
We tried lots of different ways of saying this and they were all a
little messy. Yours mixes concerns and is rather verbose, so not
convinced it is actually an improvement.
well, the way it's presently said in the spec is (to me) pretty darn hard to
understand.
which concerns are missed in the suggested reformulation?
I said "mixes" :-)
That is, it mixes the hashing up with the definition of the list.
Other than that, I'm OK with the rephrasing.
Anyone can submit certificates to certificate logs for public
auditing, however, since certificates will not be accepted by clients
unless logged, it is expected that certificate owners or their CAs
will usually submit them. A log is a single, ever-growing, append-
only Merkle Tree of such certificates.
When a valid certificate is submitted to a log, the log MUST
immediately return a Signed Certificate Timestamp (SCT). The SCT is
the log's promise to incorporate the certificate in the Merkle Tree
within a fixed amount of time known as the Maximum Merge Delay (MMD).
If the log has previously seen the certificate, it MAY return the
same SCT as it returned before.
What if the submitted end entity cert is the same, but the certificate
chain
is different (yet valid)?
The purpose of the chain is to:
a) Prevent spam, and
b) Identify who to blame in the event of a misissue.
Alternate chains presumably don't actually change the direct blame,
and so I see no reason to do other than what the I-D says - i.e.
return the same SCT as before.
ok. tho i wonder if there's any value in caching the alternate chain or the
new portions thereof.
Logs are free to keep additional information, of course :-)
We probably will.
Its not clear that its particularly useful.
TLS servers MUST present an SCT from
one or more logs to the client together with the certificate. TLS
clients MUST reject certificates that do not have a valid SCT for the
end-entity certificate.
[ see comment (7) above ]
Periodically, each log appends all its new entries to the Merkle
Tree, and signs the root of the tree. Clients and auditors can thus
Should "Clients and auditors" actually be "TLS Clients, log monitors, and
log auditors" ?
Bearing in mind that these are actually roles rather than distinct
entities, it should probably just say "auditors".
I dunno, that may loose some readers. in any case, the distinction between
roles and distinct entities should probably be mentioned/discussed e.g. in
the Introduction or thereabouts.
Yeah.
<snip>
3.1. Log Entries
Anyone can submit a certificate to any log. In order to enable
attribution of each logged certificate to its issuer, the log SHALL
publish a list of acceptable root certificates (this list might
usefully be the union of root certificates trusted by major browser
vendors). Each submitted certificate MUST be accompanied by all
additional certificates required to verify the certificate chain up
to an accepted root certificate. The root certificate itself MAY be
omitted from this list.
Alternatively, (root as well as intermediate) Certificate Authorities
Additionally? which manner is the experiment going to operate, or is it
TBD
?
Not sure what you mean? The log will accept either type of submission.
oh, sorry, I meant s/Alternatively/Additionally/
Probably the other way round? In any case, as I said, the log will
accept either type - some CAs seem to prefer one, some the other.
Structure of the Signed Certificate Timestamp:
The SCT discussion here should probably be its own subsection.
OK.
enum { certificate_timestamp(0), tree_hash(1), 255 }
SignatureType;
enum { v1(0), 255 }
Version;
struct {
opaque key_id[32];
} LogID;
opaque CtExtensions<0..2^16-1>;
"key_id" is the SHA-256 hash of the log's public key, calculated over
the DER encoding of the key represented as SubjectPublicKeyInfo.
I'd place the above paragraph regarding "key_id" down below the
SignedCertificateTimestamp definition.
struct {
Version sct_version;
LogID id;
uint64 timestamp;
CtExtensions extensions;
digitally-signed struct {
Version sct_version;
SignatureType signature_type = certificate_timestamp;
uint64 timestamp;
LogEntryType entry_type;
select(entry_type) {
case x509_entry: ASN.1Cert;
case precert_entry: ASN.1Cert;
} signed_entry;
CtExtensions extensions;
};
} SignedCertificateTimestamp;
The encoding of the digitally-signed element is defined in [RFC5246].
I would add a few words here summarizing that what happens here is that
the
digitally-signed struct here is replaced in the actual serialized binary
structure by a struct DigitallySigned and cross-ref to S4.7 of RFC5246.
Except it isn't :-)
ok, then I don't understand what's going on here. RFC5246 S4.7 sez...
A digitally-signed element is encoded as a struct DigitallySigned:
struct {
SignatureAndHashAlgorithm algorithm;
opaque signature<0..2^16-1>;
} DigitallySigned;
..and I take the "digitally-signed struct" within SignedCertificateTimestamp
to be a "digitally-signed element". Please help me understand what am I
missing?
We've actually significantly revamped this whole thing, so hopefully
you'll find it clearer in the next version.
5.1. Monitor
Monitors watch logs and check that they behave correctly. They also
watch for certificates of interest.
"Monitor" should be "Log Monitor" ?
There's no other kind of monitor :-)
in the explicit context of this spec, agreed. but in general I
favor/advocate creation and use of more fully descriptive words/phrases so
at least one is more likely to find them with a search when you're wondering
what sort of "monitor" someone down the road is yammering on about.
[ I would add a terminology section to the spec ]
5.2. Auditor
Auditors take partial information about a log as input and verify
that this information is consistent with other partial information
they have. An auditor might be an integral component of a TLS
client, it might be a standalone service or it might be a secondary
function of a monitor.
"Auditor" should be "Log Auditor" ?
And there's no other kind of auditor.
see above :)
e.g. in the overall webpki world, there /are/ other forms of auditors (e.g.
the ones noted here <http://wiki.cacert.org/Audit/CriteriaAlphabetSoup>) and
so it's worth using a more descriptive term, imv.
8. Efficiency Considerations
The Merkle tree design serves the purpose of keeping communication
overhead low.
Auditing logs for integrity does not require third parties to
maintain a copy of each entire log. The Signed Tree Heads can be
updated as new entries become available, without recomputing entire
trees. Third party auditors need only fetch the Merkle consistency
proofs against a log's existing STH to efficiently verify the append-
only property of updates to their Merkle Trees, without auditing the
entire tree.
The above could be explained in more detail, and S5.1 should be
cross-referenced. Is the last sentence above essentially a summary of
step
#8 in S5.1? Or are there differences?
It is a summary of 5.1
ok, but it'd be helpful to state that and crossref S5.1.
Actually, its the auditor not the monitor, but I have referenced it now.
wrt he JSON stuff...
1. The client messages S4 don't explicitly lay out the syntax for request
messages or responses. E.g., for S4.1 "Add Chain to Log", is the input a
stand-alone JSON text array, or a JSON text object containing a JSON text
array?
The term "JSON object" as used in the first paragraph is ambiguous and
perhaps what is mean is simply "JSON texts" or "JSON text objects or JSON
text arrays". RFC4627 clearly defines "JSON text", and should be cited.
But
RFC4627 is a little ambiguous itself regarding "JSON object" and so I
suggest these definitions:
JSON text object: A JSON text matching the "object" ABNF production
in Section 2.2 of [RFC4627].
JSON text array: A JSON text matching the "array" ABNF production
in Section 2.3 of [RFC4627].
I agree that RFC 4627 should be cited and I will correct that.
ok.
The
rest of this confuses me: JSON is a textual representation of
structured data, as it states in the RFC. It defines an object quite
clearly
" An object is an unordered collection of zero or more name/value
pairs, where a name is a string and a value is a string, number,
boolean, null, object, or array."
well, yes, but that's in just the introduction of RFC 4627.
Defining a "JSON text object" seems pointless to me - clearly a JSON
object is an object as defined by JSON, surely? Introducing another
term seems like to add confusion rather than remove it.
note that "JSON text" is very explicitly defined in RFC4627 at the beginning
of S2 as..
A JSON text is a sequence of tokens. The set of tokens includes six
structural characters, strings, numbers, and three literal names.
A JSON text is a serialized object or array.
JSON-text = object / array
the reason I flagged this issue is that I just recently reviewed a different
internet-draft where they'd confused a JSON object -- in the sense of a JSON
text matching the object production of S2.2 RFC4627 -- and a "JSON object"
in the sense of an abstract programming construct, and they didn't
understand that in the protocol on-the-wire world a JSON object /is a
string/ (i.e. it is string-serialized according to the grammar of RFC4627).
It seems the term "JSON object" is ambiguous depending on whether you're
looking at it from a programming perspective or an on-the-wire protocol
perspective (eg see <http://www.json.org/javadoc/org/json/JSONObject.html>
which talks about "internal form" and "external form" (ugh)), hence my
(perhaps feeble) attempt to invent a more explicit term for the on-the-wire
protocol form.
So maybe a more palatable definition would be..
JSON object: A JSON text matching the "object" ABNF production
in Section 2.2 of [RFC4627].
ok
?
---
end
|
|