RFR: 8257189: Handle concurrent updates of MH.form better

[Peter Levart]

On Tue, 1 Dec 2020 10:05:04 GMT, Peter Levart <plevart at openjdk.org> wrote:

>> Concurrent updates may lead to redundant LambdaForms created and unnecessary class loading when those are compiled. 
>> 
>> Most notably, it severely affects MethodHandle customization: when a MethodHandle is called from multiple threads, every thread starts customization which takes enough time for other threads to join, but only one of those customizations will be picked.
>> 
>> Coordination between threads requesting the updates and letting a single thread proceed avoids the aforementioned problem. Moreover, there's no need to wait until the update in-flight is over: all other threads (except the one performing the update) can just proceed with the invocation using the existing MH.form. 
>> 
>> Testing:
>> - manually monitored the behavior on a stress test from [JDK-8252049](https://bugs.openjdk.java.net/browse/JDK-8252049)
>> - tier1-4
>
> src/java.base/share/classes/java/lang/invoke/MethodHandle.java line 1783:
> 
>> 1781:             } finally {
>> 1782:                 updateInProgress = false;
>> 1783:             }
> 
> Line 1782. I understand that in case the try block fails to update `form` field, resetting `updateInProgress` is desirable in order for next caller to attempt to update the form next time, but in case (the majority case) the try block successfully updates the `form` field, resetting the `updateInProgress` back to `false` opens a theoretical window (very improbable) for a concurrent thread to re-execute the updater function at least. Perhaps you could write `catch (Throwable e)` instead of `finally` and only in case of failure, reset the `updateInProgress` field and re-throw the exception. WDYT?
> 
> It is also very strange that the code is updating a final `form` field using Unsafe. Is this only to ensure that reads of that field scattered around in code are accompanied with a suitable read fence if needed (on some architectures)? How does this work considering (I may be wrong, but I think I heard) that VM trusts final fields in java.lang.invoke package.

It seems that I was right. See `ciField.cpp`:

static bool trust_final_non_static_fields(ciInstanceKlass* holder) {
  if (holder == NULL)
    return false;
  if (holder->name() == ciSymbol::java_lang_System())
    // Never trust strangely unstable finals:  System.out, etc.
    return false;
  // Even if general trusting is disabled, trust system-built closures in these packages.
  if (holder->is_in_package("java/lang/invoke") || holder->is_in_package("sun/invoke") ||
      holder->is_in_package("jdk/internal/foreign") || holder->is_in_package("jdk/incubator/foreign") ||
      holder->is_in_package("jdk/internal/vm/vector") || holder->is_in_package("jdk/incubator/vector") ||
      holder->is_in_package("java/lang"))
    return true;

-------------

PR: https://git.openjdk.java.net/jdk/pull/1472