Gregory, Your are right, reported exception is held by a local reference and thus cannot be collected. So it's safe to call ExceptionClear() before NewGlobalRef().
Running Sun's VM with -Xcheck:jni just confirms your words. Invoking NewGlobalRef() before ExceptionClear() leads to error message: FATAL ERROR in native method: JNI call made with exception pending I just forgot about this useful option, thank you for remanding me. I'll fix JDWP agent to call ExceptionClear() first. Thanks. Ivan On 3/23/07, Gregory Shimansky <[EMAIL PROTECTED]> wrote:
Ivan Popov wrote: > Gregory, > > Since ExceptionClear() zeroes reference to exception object, it may be > collected before subsequent call to NewGlobalRef() can pin it. The > only safe way to prevent exception from garbage collection and be able > to report it to debugger is to call NewGlobalRef() before > ExceptionClear(). I cannot find any other solution. If you call NewGlobalRef, then you are calling it on some object reference, right? If this reference points to an exception object, it won't be collected since there is a live reference to it from your code. Even if the reference to exception object in TLS is NULL, your code holding the reference won't let the object to be collected. > Though this does not follow well JNI spec (which is quite vague in > many points), it works with RI. And I'd like NewGlobalRef() to be > fixed and don't return NULL in this case. Do you agree? Does it work in the same way with -Xcheck:jni? > On 3/23/07, Gregory Shimansky <[EMAIL PROTECTED]> wrote: >> Ivan Popov wrote: >> > Working on HARMONY-3304 [1] I noticed problem with JNI function >> > NewGlobalRef() in DRLVM. If it is called for a pending exception >> > object before ExceptionClear() is invoked, it returns NULL, which is >> > interpreted as out of memory according to JNI spec [2]. This causes >> > errors in debug support. In Sun/BEA VM non-NULL reference is returned >> > in this case. >> >> Actually JNI specification [1] states that only 3 JNI functions may be >> safely called in exception state. Looking at the code of NewGloabRef I >> see that the check for exception was added at revision 489694 as a >> result of committing HARMONY-2817. >> >> > Simple workaround is to call ExceptionClear() before NewGlobalRef(). >> > However, I'm not sure if ExceptionClear() won't dispose exception >> >> ExceptionClear zeroes a reference to exception object in TLS which is >> considered as no exception state. The object itself may be collected >> later if there are no live references to it left. >> >> > object itself. To my mind it's better to fix NewGlobalRef() and make >> > it compatible with RI. >> >> [1] >> http://java.sun.com/j2se/1.5.0/docs/guide/jni/spec/design.html#wp17626 >> >> -- >> Gregory >> >> > -- Gregory
