Re: RFR(XS): 8222005: ClassRedefinition crashes with: guarantee(false) failed: OLD and/or OBSOLETE method(s) found

[serguei.spit...@oracle.com]

Thank you a lot for review, Coleen!
Serguei


On 6/3/20 08:50, coleen.phillim...@oracle.com wrote:

Hi Serguei,
This change looks great.  Thank you for fixing this!
Coleen

On 5/28/20 7:16 PM, serguei.spit...@oracle.com wrote:

Hi Coleen,

The updated webrev version is:
  http://cr.openjdk.java.net/~sspitsyn/webrevs/2020/jvmti-redef.3/

It has your suggestions addressed:
 - remove log_is_enabled conditions
 - move ResourceMark's out of loops

Thanks,
Serguei


On 5/28/20 14:44, serguei.spit...@oracle.com wrote:

Hi Coleen,

Thank you a lot for reviewing this!


On 5/28/20 12:48, coleen.phillim...@oracle.com wrote:

Hi Serguei,
Sorry for the delay reviewing this again.

On 5/18/20 3:30 AM, serguei.spit...@oracle.com wrote:

Hi Coleen and potential reviewers,

Now, the webrev:
  http://cr.openjdk.java.net/~sspitsyn/webrevs/2020/jvmti-redef.2/

has a complete fix for all three failure modes related to the guarantee about OLD and OBSOLETE methods.

The root cause are the following optimizations:

 1) Optimization based on the flag ik->is_being_redefined():
    The problem is that the cpcache method entries of such classes are not being adjusted.
    It is explained below in the initial RFR summary.
    The fix is to get rid of this optimization.

This seems like a good thing to do even though (actually especially because) I can't re-imagine the logic that went into this optimization.

Probably, I've not explained it well enough.
The logic was that the class marked as is_being_redefined was considered as being redefined in the current redefinition operation.
For classes redefined in current redefinition the cpcache is empty, so there is  nothing to adjust.
The problem is that classes can be marked as is_being_redefined by doit_prologue of one of the following redefinition operations.
In such a case, the VM_RedefineClasses::CheckClass::do_klass fails with this guarantee.
It is because the VM_RedefineClasses::CheckClass::do_klass does not have this optimization
and does not skip such classes as the VM_RedefineClasses::AdjustAndCleanMetadata::do_class.
Without this catch this issue could have unknown consequences in the future execution far away from the root cause.

 2) Optimization for array classes based on the flag _has_redefined_Object.
    The problem is that the vtable method entries are not adjusted for array classes.
    The array classes have to be adjusted even if the java.lang.Object was redefined
    by one of previous VM_RedefineClasses operation, not only if it was redefined in
    the current VM_RedefineClasses operation. The fix is is follow this requirement.

This I can't understand.  The redefinitions are serialized in safepoints, so why would you need to replace vtable entries for arrays if java.lang.Object isn't redefined in this safepoint?

The VM_RedefineClasses::CheckClass::do_klass fails with the same guarantee because of this.
It never fails this way with this optimization relaxed.
I've already broke my head trying to understand it.
It can be because of another bug we don't know yet.

 3) Optimization based on the flag _has_null_class_loader which assumes that the Hotspot
    does not support delegation from the bootstrap class loader to a user-defined class
    loader. The assumption is that if the current class being redefined has a user-defined
    class loader as its defining class loader, then all classes loaded by the bootstrap
    class loader can be skipped for vtable/itable method entries adjustment.
    The problem is that this assumption is not really correct. There are classes that
    still need the adjustment. For instance, the class java.util.IdentityHashMap$KeyIterator
    loaded by the bootstrap class loader has the vtable/itable references to the method:
       java.util.Iterator.forEachRemaining(java.util.function.Consumer)
    The class java.util.Iterator is defined by a user-defined class loader.
    The fix is to get rid of this optimization.

Also with this optimization, I'm not sure what the logic was that determined that this was safe, so it's best to remove it.  Above makes sense.

I don't know the full theory behind this optimization. We only have a comment.

All three failure modes are observed with the -Xcomp flag.
With all three fixes above in place, the Kitchensink does not fail with this guarantee anymore.

http://cr.openjdk.java.net/~sspitsyn/webrevs/2020/jvmti-redef.2/src/hotspot/share/oops/cpCache.cpp.udiff.html

For logging, the log_trace function will also repeat the 'if' statement and not allocate the external_name() if logging isn't specified, so you don't need the 'if' statement above.

+  if (log_is_enabled(Trace, redefine, class, update)) {
+    log_trace(redefine, class, update, constantpool)
+          ("cpc %s entry update: %s", entry_type, new_method->external_name());

http://cr.openjdk.java.net/~sspitsyn/webrevs/2020/jvmti-redef.2/src/hotspot/share/oops/klassVtable.cpp.udiff.html

Same in two cases here, and you could move the ResourceMark outside the loop at the top.

Good suggestions, taken.

Thanks!
Serguei

Thanks,
Coleen

There is still a JIT compiler relted failure:
  https://bugs.openjdk.java.net/browse/JDK-8245128
    Kitchensink fails with: assert(destination == (address)-1 || destination == entry) failed: b) MT-unsafe modification of inline cache

I also saw this failure but just once:
  https://bugs.openjdk.java.net/browse/JDK-8245126
    Kitchensink fails with: assert(!method->is_old()) failed: Should not be installing old methods

Thanks,
Serguei


On 5/15/20 15:14, serguei.spit...@oracle.com wrote:

Hi Coleen,

Thanks a lot for review!
Good suggestion, will use it.

In fact, I've found two more related problems with the same guarantee.
One is with vtable method entries adjustment and another with itable.
This webrev version includes a fix for the vtable related issue:
  http://cr.openjdk.java.net/~sspitsyn/webrevs/2020/jvmti-redef.2/

I'm still investigating the itable related issue.

It is interesting that the Kitchensink with Instrumentation modules enabled is like a Pandora box full of surprises.
New problems are getting discovered after some road blocks are removed.
I've just filed a couple of compiler bugs discovered in this mode of testing:
  https://bugs.openjdk.java.net/browse/JDK-8245126
    Kitchensink fails with: assert(!method->is_old()) failed: Should not be installing old methods

  https://bugs.openjdk.java.net/browse/JDK-8245128
    Kitchensink fails with: assert(destination == (address)-1 || destination == entry) failed: b) MT-unsafe modification of inline cache


Thanks,
Serguei  


On 5/15/20 05:12, coleen.phillim...@oracle.com wrote:

Serguei,

Good find!!  The fix looks good.  I'm sure the optimization wasn't noticeable and thank you for the additional comments.

There is a Method::external_name() function that I believe prints all the things you want in the logging here:

http://cr.openjdk.java.net/~sspitsyn/webrevs/2020/jvmti-redef.1/src/hotspot/share/oops/cpCache.cpp.udiff.html

I don't need to see another webrev if you make this change.

Thanks,
Coleen

On 5/14/20 12:26 PM, serguei.spit...@oracle.com wrote:

Please, review a fix for The Kitchensink bug:
  https://bugs.openjdk.java.net/browse/JDK-8222005

Webrev:
  http://cr.openjdk.java.net/~sspitsyn/webrevs/2020/jvmti-redef.1/

Summary:
  The VM_RedefineClasses::doit() uses two helper classes to walk all VM classes.
  First is AdjustAndCleanMetadata to adjust method entries in the vtables/itables/cpcaches.
  Second is CheckClass to check that adjustments for all method entries are correct.
  The Kitchensink test is failing with two modes:
    - guarantee(false) failed: OLD and/or OBSOLETE method(s) found in the
      VM_RedefineClasses::CheckClass::do_klass()
    - SIGSEGV in the ConstantPoolCacheEntry::get_interesting_method_entry() in context
      of VM_RedefineClasses::CheckClass::do_klass() execution

  The second failure mode is rare. In is before the first one in the code path.
  The root cause of both is that the VM_RedefineClasses::AdjustAndCleanMetadata::do_klass()
  is skipping the cpcache update for classes that are being redefined assuming they are
  being redefined by the current VM_RedefineClasses operation. In such cases, the adjustment
  is not needed as the cpcache is empty. The problem is that the assumption above is wrong.
  The class can also be redefined by another VM_RedefineClasses operation which has already
  executed its doit_prologue. The cpcache djustment for such class is necessary.
  The fix is to always call the cp_cache->adjust_method_entries() even if the class is
  being redefined by the current VM_RedefineClasses operation. It is possible to skip it
  but it will add extra complexity to the code.
  The fix also includes minor tweak in the cpCache.cpp to include method's class name to
  the redefinition cpcache log.

Testing:
  Ran Kitchensink test locally on a Linux server with the Instrumentation module enabled.
  The test does not fail anymore.
  In progress, a mach5 tiers 1-5 and runs and separate mach5 Kitchensink run.

Thanks,
Serguei