RFR: 8273188: java/lang/instrument/BootClassPath/BootClassPathTest.sh fails with "FATAL ERROR in native method: processing of -javaagent failed, processJavaStart failed"
dholmes at openjdk.java.net
Wed Sep 8 23:41:57 UTC 2021
On Wed, 8 Sep 2021 22:15:12 GMT, Naoto Sato <naoto at openjdk.org> wrote:
> The gist of the issue is that the test case now always creates the boot classpath with non-ASCII chars appended, because the default encoding is fixed to UTF-8 with the fix to JDK-8260265.
> On macOS, javaagent tries to load the class with US-ASCII determined by nl_langinfo() (in unix/.../EncodingSupport_md.c). The file system encoding is always UTF-8 on mac, so no need to use the determined encoding by nl_langinfo().
> On Linux, the "defaultEncoding" in Setup.java used to be US-ASCII in "C" locale (which matches the result from nl_langinfo()), so no non-ASCII suffix was appended to the `boot` path. But now the "defaultEncoding" is always UTF-8, the setup code appends the non-ASCII suffix, which fails to read in the native code using iconv with US-ASCII. The setup code should use the encoding from "sun.jnu.encoding" instead. Actually, the code there was copied from UnicodeTest.java which was modified with JDK-8260265, so the same fix needs to be applied.
> Tier5 run succeeds with one unrelated failure. Also, fixed some typos, e.g. "Uft8" -> "Utf8"
Based on what you've said the core functional change seems okay. Good catch on the uft typo!
Pre-existing: The initialization logic in this code is quite odd for the case when no conversion is necessary (we call `utfInitialize` on every call to `convertUtf8ToPlatformString`!), but I assume we do not call `appendBootClassPath` and `appendToClassLoaderSearch` very often?
Marked as reviewed by dholmes (Reviewer).
More information about the core-libs-dev