RFR: 8349546: Linux support for Kerberos "nativeccache" functionality [v13]

[Nick Hall]

On Tue, 18 Nov 2025 12:21:09 GMT, Nick Hall <duke at openjdk.org> wrote:

>> _Purpose_
>> 
>> This PR allows Linux based applications using JAAS to acquire Kerberos TGTs natively using the local system's Kerberos libraries/configuration, building on existing support on Windows/MacOSX.
>> 
>> _Rationale_
>> 
>> Currently the (pure java) JAAS codebase only supports file-based credential caches (ccaches).  There are many other useful types of ccache accessible via the local system libraries; this change allows credentials to be acquired natively using those libraries, and thus adds support for all other ccache types supported by the local system (e.g. KCM, in-memory and kernel types),  This support already exists on MacOSX and Windows.
>> 
>> The code change here largely uses the MacOSX code, edited for Linux with associated build system changes. It also adds an appropriate jtreg test which uses some native test helper code to manufacture an in-memory cache, and then uses the new code to acquire these credentials natively.  This has been tested on Linux/Mac and the jtreg test passes on each (I couldn't see any existing tests on MacOSX for this feature).
>> 
>> Additionally this PR fixes a bug that's existed for a while (see L585-588 in `nativeccache.c`) - without this code, this is a 100% reproducible segfault on Linux (it's unclear why this hasn't affected the Mac JVMs up to now, probably just no calling code that provides an empty list of addresses).  It also fixes a (non problem) typo in the variable name in a function prototype.
>> 
>> _Implementation Detail_
>> 
>> Note that there were multiple possible ways of doing this:
>> 
>> 1) Duplicate the MacOSX `nativeccache.c`, edit lightly for Linux and build a new library on Linux only (`liblinuxkrb5`), leaving MacOSX largely unchanged, but at the expense of this code duplication.
>> 
>> 2) Create a new shared library used on both platforms with conditional compilation to manage the differences.  This necessitates a library name change on MacOSX and potentially knock-on packaging changes on that platform, which seemed a potentially expensive side-effect.
>> 
>> 3) Create a shared `nativeccache.c` (using `EXTRA_SRC` in the build) and build separate MacOSX/Linux libraries.  This allows the MacOSX library name to remain unchanged, and only adds a new library in Linux.
>> 
>> I tried all three options; 3 seemed to be the best compromise all around, although is one of the options that effectively introduces a "no-op" change on MacOSX as a result.  Hopefully the additional jtreg test is sufficient to compensat...
>
> Nick Hall has updated the pull request incrementally with one additional commit since the last revision:
> 
>   Clean up jtreg run/compile directives, attend to other review comments

There's prior art here for CUPS - this is a mandatory compile-time and run-time dependency.  I guess the difference with CUPS is that Java can't print without it, whereas JAAS _can_ auth using Kerberos without my code, it's just limited to a file ccache.

It turns out that this limitation actually works to our advantage.  The existing code will first try and use the pure Java code to acquire a file ccache using a series of hard-coded defaults:

     * 1. KRB5CCNAME (bare file name without FILE:)
     * 2. /tmp/krb5cc_<uid> on unix systems
     * 3. <user.home>/krb5cc_<user.name>
     * 4. <user.home>/krb5cc (if can't get <user.name>)

(I'm not sure of the provenance of 3 and 4, but 1 and 2 are reasonable)

It then checks that the crypto is something it can handle.

If the code successfully finds a supported ccache, it will succeed _whether the new native lib is loadable or not_.  If it does not find a supported ccache this way, it will then try and load the native lib (and potentially fail with an `UnsatisfiedLinkError`) at this point.

(FWIW, I've tested this locally by temporarily making the system libkrb5 library inaccessible, then running a test with a regular FILE: ccache, and it worked as above.)

Assuming we did build it in by default, I suspect most people using FILE: ccaches will not even get to the native library load - and anyone trying to use unsupported ccache types/crypto with an older Java version would have got an error anyway, it's just that now this might be an `UnsatisfiedLinkError` instead of a `LoginException`.

I imagine this code was written this way for similar reasons to the discussion we're having here, on whichever of Windows/MacOS this was first introduced.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/28075#issuecomment-3554366847