Hi Martin, On 12/10/2015 12:18 AM, Martin Buchholz wrote:
I finally studied ForkJoinTask.getThrowableException - I was not aware that we were doing reflection based exception cloning. In practice it probably muddles through, but it just seems too hacky to use in a java core library. We're relying on common conventions, but in principle we are invoking random code with an unknown spec. The stack traces created are fake in the sense that they are not created naturally, and the thrown exception may differ from the original in significant ways, most notably missing the message string. With heroic efforts we can make it safer, e.g. examine the bytecode for the constructors we are invoking, and check whether they call the super constructors in the right way, but that will be a major engineering project by itself.
I assume the aim of exception cloning in ForkJoinTask.getThrowableException is to provide the user with a stack-trace that originates in its own thread. This helps trace the origin of the exception in case it's handled way up in the call stack. In addition, the original stack trace is also preserved (original exception is added as a cause).
This case does not suffer from the visibility of original exception before ForkJoinTask.getThrowableException is called, so perhaps the aim could be achieved in a slightly different manner: just return the original exception, but add a newly constructed CrossThreadPropagationException as a suppressed exception to it. Printing such augmented original exception would then reveal both stack traces. What do you think?
For CompletableFuture.whenComplete: I also keep moving in the direction of less magic. It is perfectly possible for a well-written whenComplete action to catch all Throwables, add the source exception as a suppressed exception, and rethrow. In fact, perhaps people have already tried this and discovered we incorrectly throw the source exception. I think we should really accept our mistake and switch to throwing the action exception rather than the source exception. Perhaps we should only call addSuppressed if the source exception is not already in the list of suppressed exceptions.
Less is more. I would even not bother with the additional logic of conditionally adding source exception to the list of suppressed exceptions. I doubt many codes out there do that in the whenComplete handler since such exception was ignored up until now. And if anybody does that, she will get the source exception listed twice - not a big deal. If javadoc describes the behavior, this will not happen frequently.
Regards, Peter
On Thu, Dec 3, 2015 at 9:54 AM, Jason Mehrens <jason_mehr...@hotmail.com> wrote:Hi Peter, I've done this trick before to perform Throwable cloning. You have to hunt for the constructors in this order: 1. Walk the type of the cause and super types by calling getConstructor(String, type of cause). (java.io.WriteAbortedException and javax.mail.MessagingException) 2. Walk the type of the cause and super types by calling getConstructor(type of cause, String) (I've seen this but, I can't seem to find an example) 3. getConstructor(String) + initCause (java.io.InvalidObjectException) 4. Walk the type of the cause and super types by calling getConstructor(type of cause). (java.awt.print.PrinterIOException) 5. getConstructor() + initCause. (java.nio.BufferOverflowException) 6. Look at the return type of methods declared in the cause type and assume there is a constructor matching the type or no repeating types. (java.nio.charset.MalformedInputException, java.lang.EnumConstantNotPresentException, java.util.regex.PatternSyntaxException) In the end all of that can still fail. Exceptions can be private subclasses of public classes so that gets interesting with reflection too. Jason =========== What about the 4th option (keep current behavior, but try the best-effort with reflection to make new exception of the same type): ... } catch (Throwable ex) { if (x == null) { x = ex; } else { // try to create new exception with same: // type, cause, suppressed exceptions and stack-trace... Throwable nx; Class<?> xClass = x.getClass(); try { Constructor<?> xConstr = xClass.getConstructor(Throwable.class); nx = (Throwable) xConstr.newInstance(x.getCause()); } catch (Exception e1) { try { nx = (Throwable) xClass.newInstance(); nx.initCause(x.getCause()); } catch (Exception e2) { // no luck nx = null; } } if (nx != null) { // inherit stack-trace of original exception nx.setStackTrace(x.getStackTrace()); // inherit suppressed exceptions for (Throwable sx : x.getSuppressed()) { nx.addSuppressed(sx); } // add 'ex' as a final suppressed exception nx.addSuppressed(ex); x = nx; } } } completeThrowable(x, r); ... Note that such code and similar code in ForkJoinTask.getThrowableException will probably have to be modified for jigsaw to include dynamic addition of read-edge to the module of exception class... Regards, Peter
