I don't think we need it. We're already getting a similar message as
things stand now. The existing message doesn't correlate the extension
with the message, but you can figure out which message is being
processed by looking at where it falls in the trace.
I'll push this as-is. Thanks for the review and the improvements to the
overall fix.
--Jamil
On 1/5/2020 6:45 PM, Xuelei Fan wrote:
Hi Jamil,
Thanks for the new update, which looks pretty good to me.
If have something (like SSLLogger.debug()) for line 96 could mitigate
your worries about the debug log?
SSLExtensions.java:
- 96 // debug log to ignore unknown extension for handshakeType
+ if (SSLLogger.isOn && ...) {
+ SSLLogger.debug("Ignore unknown or unsupported extension (" +
+ SSLExtension.nameOf(extId) + ") in " +
+ SSLHandshake.nameOf(handshakeType));
+ }
Thanks,
Xuelei
On 1/5/2020 2:20 PM, Jamil Nimeh wrote:
Hi Xuelei, et al.,
I have a new version of the webrev here:
https://cr.openjdk.java.net/~jnimeh/reviews/8236039/webrev.02
I thought a little more about what you said about consumers and tried
looking at the logging both with and without the CR consumer in
place. With your changes to the SSLExtensions constructor, combined
with the new SSLExtension.nameOf(int), I think we can not have the
consumer and the resulting logs look good. You still get a warning
when processing a CR status_request, but now an "unknown or
unsupported" message that identifies "status_request (5)" makes more
sense. The status_request extension is identified by name now but
unsupported in the CR message which is accurate, but no consumer
overhead at all, debug or not. Best of both worlds.
When we get to doing the client-side OCSP stapling feature, then
we'll need server side producers and client-side consumers. But all
that fun is for a later date.
Thanks,
--Jamil
On 1/4/20 6:41 PM, Xuelei Fan wrote:
It's a good idea to me to have a SSLExtension.nameOf(int) method.
Thanks,
Xuelei
On 1/4/2020 3:53 PM, Jamil Nimeh wrote:
Let me throw out another idea. I think the actual parsing only
happens in the consumer if you call the spec constructor. You
could have a consumer registered that simply returns null. That
avoids the extra unknown/unsupported log message that we don't need
and you don't have the cost of actually parsing the extension. It
should just skip over it. The only performance hit you take is the
actual call into one additional method. That shouldn't have a
major performance impact.
Just to see how it would work, I created SSLExtension.nameOf(int)
which takes the extension ID and returns a name or "unknown
extension". I did that because I didn't want to mess with
valueOf(SSLHandshake, int) since there's other logic depending on
that returning null for extensions we don't support.
The result was that the unsupported-but-known extensions have their
names presented, which is a nice thing for those looking at the log.
--Jamil
On 1/4/20 3:35 PM, Xuelei Fan wrote:
We may not want to parse the extension if the debug is not
enabled. It may not worthy to define consume for debug only.
How about replace “unknown” with “unsupported”? And parse the
extension id to literal name if the name is known?
Xuelei
On Jan 4, 2020, at 3:19 PM, Jamil Nimeh
<[email protected]> wrote:
Hi Xuelei, I backed out my change and went solely with the
approach you suggested below. It works in that it allows the
handshake to proceed. I notice these debug log messages though
during consumption of the CR message from the server:
javax.net.ssl|DEBUG|01|main|2020-01-04 15:02:26.636
PST|SSLExtensions.java:135|Ignore unknown or unsupported extension (
"unknown extension (5)": {
}
...
javax.net.ssl|DEBUG|01|main|2020-01-04 15:02:26.656
PST|CertificateRequest.java:926|Consuming CertificateRequest
handshake message (
"CertificateRequest": {
"certificate_request_context": "",
"extensions": [
"unknown extension (5)": {
},
"unknown extension (18)": {
},
"signature_algorithms (13)": {
"signature schemes": [rsa_pss_rsae_sha256,
ecdsa_secp256r1_sha256, ed25519, rsa_pss_rsae_sha384,
rsa_pss_rsae_sha512, rsa_pkcs1_sha256, rsa_pkcs1_sha384,
rsa_pkcs1_sha512, ecdsa_secp384r1_sha384, ecdsa_secp521r1_sha512,
rsa_pkcs1_sha1, ecdsa_sha1]
},
"unknown extension (47)": {
0000: 00 29 00 27 30 25 31 10 30 0E 06 03 55 04 0A 0C
.).'0%1.0...U...
0010: 07 54 65 73 74 50 4B 49 31 11 30 0F 06 03 55 04
.TestPKI1.0...U.
0020: 03 0C 08 54 65 73 74 52 6F 6F 74 ...TestRoot
}
]
}
)
In the debug logs, it seems like we shouldn't call a
status_request extension an "unknown extension" because we know
what it is and can parse it. Having the entry in SSLExtension
for CR_STATUS_REQUEST and a simple do-nothing consumer gets rid
of the status_request/unknown messages and has it render in the
printing of the CR message properly.
While we're talking about the debug logs we have other extensions
that we know about and have entries in SSLExtension, but they
have no per-message registration. You can see in the CR above
that 18 (signed_certificate_timestamp) and 47
(certificate_authorities) are extensions that we were aware of
enough to put basic entries in SSLExtension for, but just didn't
put full consumer/producer support in there.
It might be nice to use the display approach that we have for an
unsupported extension for those extensions we at least know
about, but rather than say "unknown extension" at the start of
the printing at least give the name. Let me see if I can make
that work without it being too invasive.
--Jamil
On 1/4/20 9:27 AM, Xuelei Fan wrote:
CertStatusExtension.java
------------------------
1239 // The consuming happens in server side only.
typo? server -> client
It wold be nice to add a debug log that this extension get
ignored. But may not need to define this consumer as it is not
supported.
SSLExtension.java
-----------------
As this fix does not implement this feature yet, is it possible
to just define it without the on-load consumer?
Otherwise, it looks fine to me.
BTW, this issue reminders me a common problem: if an extension
is supported in a handshake message, we need to know all other
handshake messages that could use the extension, and define an
SSLExtension enum for them. Otherwise, a similar issue could
happen. I think it would be challenge to know that in practice.
So I was just wondering, could we just release the check in the
SSLExtensions.java implementation (from line 71 to 95)? If the
extension for the specific handshake type is not defined, just
ignore it, except for ServerHello? I think the behavior
complies to the TLS 1.3 protocol.
if (SSLExtension.isConsumable(extId) &&
SSLExtension.valueOf(handshakeType, extId) == null) {
...
- } else {
+ } else if (handshakeType == SSLHandshake.SERVER_HELLO) {
throw hm.handshakeContext.conContext.fatal(
Alert.UNSUPPORTED_EXTENSION,
"extension (" + extId +
") should not be presented in " +
handshakeType.name);
+ } else {
+ isSupported = false;
+ // debug log to ignore unknown extension for handshakeType
}
}
Xuelei
On 1/3/2020 10:06 AM, Jamil Nimeh wrote:
Hi All, the golang folks have been running into an issue where
our JSSE client treats the status_request extension in a
CertificateRequest message from a golang server as an unknown
extension and alerts. This quick fix will allow the client to
read and accept the extension and proceed. I believe you need
golang 1.13.x to see this take place.
This fix does not implement client-side OCSP stapling. That
will be an RFE for another day.
Bug: https://bugs.openjdk.java.net/browse/JDK-8236039
Webrev:
https://cr.openjdk.java.net/~jnimeh/reviews/8236039/webrev.01/
--Jamil
