Andrey Yarovoy created HDDS-17781:
-------------------------------------
Summary: Shaded Ozone filesystem client registers its relocated
Bouncy Castle provider as the JVM-wide "BC", breaking the application's own
Bouncy Castle
Key: HDDS-17781
URL: https://issues.apache.org/jira/browse/HDDS-17781
Project: Apache Ozone
Issue Type: Bug
Components: Ozone Client
Reporter: Andrey Yarovoy
h3. Problem
The shaded filesystem jars relocate Bouncy Castle to
\{{org.apache.ozone.shaded.org.bouncycastle}}, but the relocated provider still
calls itself \{{"BC"}} (\{{BouncyCastleProvider.PROVIDER_NAME}} is a string,
which relocation doesn't change). \{{SecurityConfig.initSecurityProvider}}
registers it with \{{Security.addProvider(new BouncyCastleProvider())}}, which
puts it in the JVM-wide provider list under the name \{{"BC"}}. Only one
provider can hold that name, and the first one registered wins.
Up to 2.2.x, the \{{SecurityConfig}} constructor does this. Every \{{ofs}} /
\{{o3fs}} initialization creates a \{{SecurityConfig}}
(\{{BasicRootedOzoneClientAdapterImpl}}, \{{BasicOzoneClientAdapterImpl}}),
with or without security enabled. So any JVM that opens an Ozone file system
registers Ozone's private, relocated BC as the JVM's \{{"BC"}}. After that:
* the application's own \{{Security.addProvider(new BouncyCastleProvider())}}
returns \{{-1}} and is silently ignored;
* \{{getInstance(..., "BC")}} in the application returns relocated objects. A
cast to a Bouncy Castle type (\{{org.bouncycastle.jce.interfaces.ECPublicKey}},
\{{BCECPublicKey}}, …) throws \{{ClassCastException}};
* the application runs on Ozone's Bouncy Castle version, not its own, without
knowing it.
If the application registers BC first, it's the other way round: Ozone silently
uses the application's BC, whatever version that is.
h3. Reproduced
{\{new SecurityConfig(new OzoneConfiguration())}}, which is what an \{{ofs}}
initialize does, then the application registers BC 1.86:
||filesystem jar||\{{"BC"}} after \{{new SecurityConfig(conf)}}||application's
\{{addProvider(BC)}}||application's cast of an EC key from \{{"BC"}}||
|2.2.0|\{{org.apache.ozone.shaded.org.bouncycastle.jce.provider.BouncyCastleProvider}}|\{{-1}},
ignored|fails: the key is \{{org.apache.ozone.shaded...BCECPublicKey}}|
|master (with HDDS-17353)|not registered|position 14|OK|
On master, HDDS-17353 removed the registration from the constructor. In the
2.3.0-SNAPSHOT jar, only \{{CertificateCodec.getCertFactory()}} (PEM
certificate paths) and \{{SecurityConfig.getProvider()}} (via
\{{HDDSKeyGenerator}}) still register BC. A bytecode scan of the jar finds no
caller of either from client code. The bug is therefore gone for today's client
paths, but the mechanism is still there:
* calling \{{SecurityConfig.initSecurityProvider("BC")}} with the
2.3.0-SNAPSHOT jar still registers the relocated provider globally, with the
same effects as above;
* any future client path that reaches a by-name BC lookup brings the bug back,
and nothing would catch it.
h3. Proposed change
# *Never register a relocated provider globally.* In
\{{SecurityConfig.initSecurityProvider}}, if
\{{BouncyCastleProvider.class.getName()}} is not
\{{org.bouncycastle.jce.provider.BouncyCastleProvider}} (i.e. this is the
shaded jar), don't touch \{{java.security.Security}}. Keep a private
\{{BouncyCastleProvider}} instance and resolve \{{provider}} to it. With
unrelocated BC (servers, the unshaded client) the behaviour doesn't change; the
server code passes the provider name to bcpkix builders
(\{{setProvider("BC")}}) and needs the global registration.
# *Look up by instance, not by name, in code shared with the client:*
#* \{{CertificateCodec.getCertFactory()}}:
\{{CertificateFactory.getInstance("X.509", provider)}};
#* \{{HDDSKeyGenerator}}: \{{KeyPairGenerator.getInstance(algo, provider)}}.
#* Both work the same for the unrelocated case.
# *Regression test against the shaded jar:* initialize an \{{ofs}} file system
(secure and non-secure configuration) and call
\{{CertificateCodec.getCertFactory()}}. Check that
\{{Security.getProvider("BC")}} is still \{{null}}, or not a relocated class,
and that an application's \{{Security.addProvider(new BouncyCastleProvider())}}
afterwards returns a position, not \{{-1}}.
# *Release lines:* for 2.2.x, the minimal fix is the constructor part of
HDDS-17353, which stops \{{new SecurityConfig(conf)}} from registering BC.
Steps 1 and 2 make it impossible in any version.
*Alternative considered:* stop relocating Bouncy Castle (HDDS-13231). That
removes the class mismatch, but the application and Ozone would then share one
BC on the classpath, with version-skew risks of its own, and it's a much bigger
change. This fix is independent of it.
h3. Compatibility
* *Applications using the shaded filesystem jar:* their own \{{"BC"}} works
again, whatever the order of initialization.
* *Ozone in the shaded jar:* uses its own BC instance for the few BC operations
it does, never the application's.
* *Servers and unshaded clients:* no change.
* No wire, proto, config or on-disk changes.
h3. Testing
* The regression test of step 3, run against \{{ozone-filesystem-hadoop3}} (and
\{{-hadoop2}}).
* *Unit, for \{{SecurityConfig}}:* with unrelocated BC,
\{{initSecurityProvider("BC")}} still registers \{{"BC"}}, and an existing
\{{"BC"}} is still reused.
* *Existing security tests:* certificate client, CA server, gRPC TLS. They
cover the unrelocated path.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]