To answer my own question the trick swt uses is to encode the build-id into the dll/so so it never needs to delete a file
See http://git.eclipse.org/c/platform/eclipse.platform.swt.git/tree/bundles/org.eclipse.swt/Eclipse%20SWT%20PI/common/org/eclipse/swt/internal/Library.java#n144 Tom Von meinem iPhone gesendet > Am 17.05.2018 um 21:50 schrieb Tom Schindl <tom.schi...@bestsolution.at>: > > Did you check how SWT does addresses the problem? If you run it outside of > OSGi it needs to deal with this problem as well. > > Tom > > Von meinem iPhone gesendet > >> Am 17.05.2018 um 21:25 schrieb Johan Vos <johan....@gluonhq.com>: >> >> Note: This is different from what is discussed in >> http://mail.openjdk.java.net/pipermail/openjfx-dev/2018-April/021762.html >> >> The following text is about bundling native code within a jar, and doesn't >> talk about how to deal with platform-specific libraries in generic jar >> files. >> >> At this moment, the standalone SDK for JavaFX 11-ea works fine, and javafx >> apps can be executed since the javafx jars will load the required native >> code from the libs that are part of the SDK. >> >> When we want to download the javafx modules as modules (or jars with >> module-info), we don't have a local SDK with native libraries. The native >> libraries need to come with the downloaded jar. There are a number of >> options to do this, and a number of ongoing discussions (e.g. see >> http://mail.openjdk.java.net/pipermail/panama-dev/2018-April/001541.html). >> In summary, jmods already have support for native code but they are not >> scoped at runtime, and jars don't have a standard way for supporting native >> code. >> >> I hope there will be a long-term solution for this in the VM in the future, >> but clearly we can't rely on this for the 11 release. >> Hence, we need to provide some JavaFX-specific fixes. >> Just to try the mechanism, I created a quick (and dirty) workaround in the >> NativeLibLoader. I committed it to a nativelibs branch of my openjfx fork: >> https://github.com/johanvos/openjdk-jfx/commit/f2c4c57a6197634e9c3bc1298ea14700057155a3 >> >> The good thing is that all JavaFX native lib loading is already delegated >> to the NativeLibLoader, so I only had to patch one file. >> >> I modified the generated javafx.graphics.jar to include all libs that are >> generated (something that can be automated in a gradle task), and could >> successfully run javafx apps using >> java -p /path/to/my/jars --add-modules javafx.controls >> without specifying java.library.path or native libs or so. >> >> This works for jars that I put myself on the classpath, hence it should >> also work for jars downloaded from maven central etc. >> >> In case there will be a standard for bundling native libs with jar files in >> a future JVM, we can remove this hack. But even in the meantime, there are >> better solutions. >> One of the best options I think we should consider is using JavaCPP ( >> https://github.com/bytedeco/javacpp) . The main author of JavaCPP, Samuel >> Audet (in cc) describes here ( >> https://github.com/tensorflow/tensorflow/issues/18397#issuecomment-381371636) >> why my hack will fail on Windows. >> >> Another advantage of JavaCPP is that it is proven to work on ios and >> Android as well. Actually, that would allow us to get rid of the >> ios-specific code that is currently in NativeLibLoader.java. >> >> Again, this is only tackling the issue of bunding native code with jars and >> not dealing with platform-specific resolution (although JavaCPP can help >> there as well). >> >> - Johan >