On 29/11/2019 07:32, Thomas Stüfe wrote:
Just read Matthias reply:
We call jplis_assert() if allocation fails. Looking at
src/java.instrument/share/native/libinstrument/JPLISAssert.h
I see that these assertions seem to be turned on all the time:
45 #define JPLISASSERT_ENABLEASSERTIONS (1)
and lands us in JPLISAssertCondition() (possible improvement here is
to evaluate the condition before the call):
58 #define jplis_assert(x) JPLISAssertCondition((jboolean)(x), #x,
THIS_FILE, __LINE__)
However, JPLISAssertCondition() is not an assert - name is misleading
- but just a printf():
39 void
40 JPLISAssertCondition( jboolean condition,
41 const char * assertionText,
42 const char * file,
43 int line) {
44 if ( !condition ) {
45 fprintf(stderr, "*** java.lang.instrument ASSERTION FAILED
***: \"%s\" at %s line: %d\n",
46 assertionText,
47 file,
48 line);
49 }
50 }
Maybe I miss something but I do not see an abort.
There is technical debt that dates back to the original development of
the JPLIS agent in JDK 5. If we run out of memory during VM startup then
doing a graceful abort seems right, pointless trying to continue. I
don't know how far you want to go with the current patch but a bit icky
to continue (even with a warning) with the wrong configuration. The late
binding agent case is different of course, I think the native method has
to complete with a pending exception rather than aborting the VM.
-Alan
