Alban Crequy wrote:
Le Sun, 15 Aug 2010 23:05:49 +0100,
Martin Morrison <[email protected]> a écrit :
I've recently been implementing XEP-0115 Entity Capabilities, both the
latest and the legacy versions, and have a few issues that I feel
should be clarified in the latest version of the spec:
1. The legacy version of the spec explicitly says (in 4.2, below
example 8):
[...]
2. The spec only uses the word "SHOULD" when specifying how the Disco
'node' attribute is formed. A receiving entity that supports both
Entity Capabilities and has multiple disco#items nodes thus has
somewhat of a dilemma in deciding how to respond to a disco#info
request for an unknown node. Should it return an <item-not-found/>
error, or assume that the remote entity has used some other mechanism
to construct the 'node' attribute in the request, and return the base
capabilities as if the node was empty?
I think <item-not-found/> is best: sending an empty capability in this
case is bad because the hash will not match and the client will discard
the reply.
This assumes that the receiving entity re-extracts the hash from the
node on reply. If I write a client that chooses to generate uuids for
the node attribute, then I am fully compliant with XEP-0115, but your
client will never tell me your capabilities.
If what you've suggested is the recommended behaviour, shouldn't the
specification use "MUST" when specifying "node#hash" as the value of the
'node' attribute?
3. Related to item 2, the following race condition can occur:
- [email protected] sends Presence to [email protected] with
an Entity Capabilities hash
- In response, juliet sends a disco#info request with the "node#hash"
as the 'node' attribute
- Meanwhile, romeo changes the feature set of his client (e.g. turns
on his camera)
- Upon receiving the disco#info request, what does romeo do?
As the 'node' attribute has been formed using the recommended method,
Romeo can establish that the hash doesn't match his current
capabilities. Should he return an error, or ignore the contents of the
'node' attribute completely and just return his current capabilities
(which will be accepted, since he will already have pushed an updated
hash via Presence)? Either way, I think it would help if the spec
specified what was expected.
For this race, it was suggested to just send <item-not-found/> in this
thread: http://mail.jabber.org/pipermail/standards/2008-May/018713.html
Sending <item-not-found/> will work: Juliet will receive the new hash
a bit later and she will send a new disco request with the new hash.
Alternatively, Telepathy-Gabble keeps a cache mapping
hash->capabilities. So with this implementation, Romeo will send the
*previous* capabilities corresponding with the requested hash. Juliet
will ask the capabilities again anyway when she receives the new hash
if she wants to know the new capabilities.
This requires Romeo to keep track of the hashes of every combination of
capabilities he has every published. It also doesn't sit well IMO with
the existence of multiple hashing algorithms - the "node#hash" 'node'
attribute doesn't tell you what the hashing algorithm is, making it even
harder to know what capabilities to return (albeit on the very unlikely
off-chance that two different capabilities hash to the same string under
two different supported algorithms).
Disco queries in XEP-0115 are not "what are your current caps?" but
"what are the caps corresponding to this hash?". It is up to the
clients to keep track of hashes and to make new disco requests when
they receive a new hash if they want to know the capabilities.
If they really are, as you say, "what are the caps corresponding to this
hash?" then shouldn't the request mirror that, by explicitly specifying
the hash algorithm and hash string? Currently the request on-the-wire
would be interpreted as "what are the caps of the node known as
'node#hash'?", with nothing to differentiate this from a plain XEP-0030
request.
I agree with the sentiments of your points, but the current
specification feels like its trying to pander to the legacy format too
much, resulting in multiple ambiguities overall.
Cheers,
Martin
Alban
