Peter, David, I updated the PR to focus on solving the problem I originally intended to solve that is observed in the wild - a deadlock when two different libraries are being loaded. The case when the same library JNI_OnLoad has FindClass with a circular dependency on loading the same library is artificial. I thought of two ways to resolve it with the following conclusions:
- not holding a lock when JNI_OnLoad is executed can only be done in a such a way that would lead to a possibility to get JNI_OnLoad executed multiple times, and the JNI code can rely on the fact that it's called only once (though the JNI spec does not say so). Decoupling dlopen/LoadLibrary from JNI_OnLoad does not solve this problem as well. - throwing something like LibraryBeingLoadedError similar to ClassCircularityError for JNI calls trying to use a class which initializes the same library is also not a good option, as it is behavior change. I as a user would not know what to do with that (remember it can appear for a different thread that just happened to be initializing the same library as well). Such a thread should just sit and wait on the lock. Both such solutions would make innocent code suffer, so are a no-go. I'm open to ideas how to resolve it, if at all, but IMO it has a different priority compared to the original bug report. I thus focused on solving the original problem. Thanks, -Aleksei On 21/05/2021 21:10, Peter Levart wrote: > > On 21/05/2021 10:29, Peter Levart wrote: >> I still haven't found a scenario of a possible deadlock when Sergei's >> initial patch is combined with... > > > Sorry Aleksei, I renamed you to Sergei without intention. Please > excuse me. > > > Peter > >
