On Fri, 19 May 2023 11:32:03 GMT, Daniel Fuchs <dfu...@openjdk.org> wrote:
>> [JDK-6956385](https://bugs.openjdk.org/browse/JDK-6956385): >> `JarURLConnection` properly tracks any `InputStream` it itself opened, and >> correspondingly closes the `JarFile` if necessary (when caches are >> disabled). However if its underlying `FileURLConnection` was used to >> retrieve a header field, that would have caused a `FileInputStream` to be >> opened which never gets closed until it is garbage collected. This means >> that an application which calls certain methods on >> `jar:file:/…something.jar!/…` URLs will leak file handles, even if >> `URLConnection` caches are supposed to be turned off. This can delay release >> of system resources, and on Windows can prevent the JAR file from being >> deleted even after it is no longer in use (for example after >> `URLClassLoader.close`). >> >> [JDK-8224095](https://bugs.openjdk.org/browse/JDK-8224095) was marked as a >> duplicate, but I think incorrectly. It refers to `FileURLConnection`, and >> seems to be complaining about the confusing API design of `URLConnection` >> generally: that it is an object which you might expect to be `Closeable` but >> in fact it is its `inputStream` which must be `close`d. In JDK-6956385, even >> when the caller _does_ specifically call `InputStream.close`, a file handle >> may be leaked. >> >> I managed to build the JDK on both Linux and (Cygwin) Windows to confirm the >> fix via >> >> >> ./build/…-x86_64-server-release/jdk/bin/java >> test/jdk/sun/net/www/protocol/jar/FileURLConnectionLeak.java >> >> >> I also ran jtreg on Linux (this test and various others in nearby packages); >> on Windows the `make test` target just hung for some reason. >> >> I marked the test `othervm` out of caution, since it is mutating global >> state, and if it fails will leak a handle and prevent scratch dir cleanup on >> Windows. >> >> (This is my first contribution, at least after the move to GitHub, so let me >> know if something is missing here technically or stylistically. None of the >> contribution guides appear to be up to date.) > > src/java.base/share/classes/sun/net/www/protocol/file/FileURLConnection.java > line 96: > >> 94: return connected; >> 95: } >> 96: > > I wonder if it would be better to add a > > > void closeInputStream() { ... } // (or void close() { ... }?) > > > that would simply close the input stream if it's not null... `close()` would be natural enough, and would encapsulate behavior a bit better. Should it then actually implement `Closeable`, permitting the implicit request in https://github.com/openjdk/jdk/pull/12871#discussion_r1198873512 to be addressed? That would not technically be an API change, since the implementation class is not exported, though it could be a behavioral change in case anyone is checking `instanceof Closeable` (or `instanceof AutoCloseable`); and would be a first step along the way to JDK-8224095, which may or may not be appropriate for something labeled a bug fix. ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/12871#discussion_r1198932229
