RFR: 8299329: Assertion failure with fastdebug build when trying to use CDS without classpath

[Ioi Lam]

On Fri, 23 Dec 2022 18:12:38 GMT, Severin Gehwolf <sgehwolf at openjdk.org> wrote:

>> It appears longest_common_app_classpath_prefix_len() is not returning correct value when there is no file separator in the path being searched backwards. Instead of returning 0 it return 1.
>> 
>> In case of empty classpath, it can result in assertion failure in check_paths():
>> 
>> `assert(strlen(rp_array->at(i)) > (size_t)runtime_prefix_len, "sanity");`
>> 
>> It can also result in incorrectly validating the app classpaths if they only differ by first character. Eg:
>> 
>> Dump time:
>>   -cp hello.jar
>> Run time:
>>   -cp mello.jar
>> 
>> This would not result in classpath mismatch!
>> 
>> This fix updates longest_common_app_classpath_prefix_len() to return 0 if no file separator character is found.
>> 
>> Signed-off-by: Ashutosh Mehra <asmehra at redhat.com>
>
> Could we add a test for this a la `test/hotspot/jtreg/runtime/cds/appcds/WrongClasspath.java`?

> @jerboaa I see WrongClasspath.java already has a test for "no classpath" condition, but it is not failing with debug build. The command line used for running "no classpaht" test reveals it is actually running with a classpath:
> 
> ```
> /work/jdk/build/base-fd/images/jdk/bin/java -cp /work/jdk/JTwork/classes/runtime/cds/appcds/WrongClasspath.d:/work/jdk/test/hotspot/jtreg/runtime/cds/appcds:/work/jdk/JTwork/classes/test/lib:/work/jdk/test/lib:/work/jtreg/build/images/jtreg/lib/javatest.jar:/work/jtreg/build/images/jtreg/lib/jtreg.jar -Xshare:on -showversion -XX:SharedArchiveFile=/work/jdk/JTwork/scratch/appcds-15h18m10s791.jsa -Dtest.timeout.factor=1.0 -XX:+UnlockDiagnosticVMOptions -XX:VerifyArchivedFields=1 -Xlog:cds Hello
> ```
> 
> The classpath is appended by jdk.test.lib.process.ProcessTools#createJavaProcessBuilder():
> 
> ```
>         args.add("-cp");
>         args.add(System.getProperty("java.class.path"));
> ```
> 
> That means the "no classpath" test is not really testing what it is expected to.

Maybe we can do this?


    // Then try to execute the archive without -classpath -- it should fail
    TestCommon.run(
-       /* "-cp", appJar, */ // <- uncomment this and the execution should succeed
+       // Force classpath to be the empty string. Note: -cp "" doesn't work on Windows.
+       "-Djava.class.path=",
        "-Xlog:cds",
        "Hello")
        .assertAbnormalExit(unableToUseMsg, mismatchMsg, hintMsg);


That should override the `-cp` parameter given by `createJavaProcessBuilder`, which appears earlier in the command-line.

Also, I think you should add the `-cp hello.jar` vs `-cp mello.jar` test case into WrongClasspath.java.

Thanks!

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

PR: https://git.openjdk.org/jdk/pull/11781