RFR (S): 8194309: JNI handle allocation failure not reported correctly

[David Holmes]

On 23/07/2020 11:54 pm, coleen.phillimore at oracle.com wrote:
> On 7/23/20 9:39 AM, David Holmes wrote:
>> On 23/07/2020 10:46 pm, coleen.phillimore at oracle.com wrote:
>>>
>>> http://cr.openjdk.java.net/~dholmes/8194309/webrev.with-test-hooks/src/hotspot/share/runtime/jniHandles.cpp.udiff.html 
>>>
>>>
>>> These functions shouldn't use UseNewCode which is false by default. 
>>> UseNewCode is for testing experimental features locally, not for 
>>> checked in code.  You should use logging for this.
>>
>> That code is not being checked in. It was purely for my testing purposes.
> 
> I really need to start reading to the bottom of the mail, I clicked 
> first.  The real version looks good to me.

Thanks for the review Coleen!

David

> Coleen
> 
>>
>> David
>> -----
>>
>>> Coleen
>>>
>>> On 7/23/20 4:52 AM, David Holmes wrote:
>>>> webrev: http://cr.openjdk.java.net/~dholmes/8194309/webrev/
>>>> bug: https://bugs.openjdk.java.net/browse/JDK-8194309
>>>>
>>>> The JNI specification states that NewLocalref and NewGloalRef return 
>>>> NULL on out-of-memory, whilst NewWeakGlobalRef throws 
>>>> OutofMemoryError. But the hotspot implementation will abort on 
>>>> out-of-memory in all cases.
>>>>
>>>> The oop storage code for global and weak-global handles already 
>>>> supports taking an AllocFailStrategy, so we simply pass the 
>>>> RETURN_NULL strategy through - and in the weak case throw a newly 
>>>> defined OutOfMemoryError for "C heap space" (as a corollary for 
>>>> "Java heap space").
>>>>
>>>> For NewLocalRef we pass the strategy through to 
>>>> JNIHandleBlock::allocate_block, where we explicitly use "new 
>>>> (std::nothrow)" to get a NULL on out-of-memory, so that we can pass 
>>>> it back.
>>>>
>>>> There are three internal calls to NewGlobalRef in jni.cpp that 
>>>> needed a NULL check added to support the new behaviour.
>>>>
>>>> In reality we know that if any of these things actually return NULL 
>>>> because C-heap is exhausted then chances are we are going to abort 
>>>> soon in any case. But to be spec compliant we make the changes.
>>>>
>>>> Note that I deliberately do not change any of the internal 
>>>> JNIHandle::make_local calls (contained in the majority of JNI 
>>>> methods) to get NULL on out-of-memory. This is because none of those 
>>>> APIs are specified in a way that even considers what should happen 
>>>> if an internal request to create a local-ref fails - so we can 
>>>> neither return NULL nor throw an exception in general. All this fix 
>>>> addresses are the three specific JNI entry points themselves.
>>>>
>>>> Also note there is no attempt with this changeset to add NULL checks 
>>>> to all the JNI code in the other JDK libraries that uses these 
>>>> API's. Interestingly quite a number already include the NULL checks.
>>>>
>>>> Testing:
>>>>
>>>> There is no practical way to test this for real so I had to use 
>>>> fault-injection. A version of the webrev with the fault-injection 
>>>> hooks and a test case, is presented here (for the record):
>>>>
>>>> http://cr.openjdk.java.net/~dholmes/8194309/webrev.with-test-hooks/
>>>>
>>>> The test is constructed so that only the JNI calls in the test 
>>>> should possibly encounter the NULL returns.
>>>>
>>>> Otherwise only sanity testing of tiers 1-3.
>>>>
>>>> Thanks,
>>>> David
>>>
>