RVM

I looked at half a dozen; they all have this as the stacktrace:

JikesRVM: internal error: recursive use of hardware exception registers (exiting)

– Stack –
at [0x684f39f8] Lorg/jikesrvm/runtime/RuntimeEntrypoints; deliverHardwareException(II)V at line 694
at [0x684f3a0c] <hardware trap>
at [0x684fb8c4] Lorg/jikesrvm/scheduler/RVMThread; <init>([BLjava/lang/Thread;Ljava/lang/String;ZZI)V
at [0x684fb914] Lorg/jikesrvm/scheduler/greenthreads/GreenThread; <init>([BLjava/lang/Thread;Ljava/lang/String;ZZI)V at line 134
at [0x684fb914] Lorg/jikesrvm/scheduler/greenthreads/GreenThread; <init>(Ljava/lang/String;)V at line 99
at [0x684fb914] Lorg/jikesrvm/scheduler/Scheduler$ThreadModel; <init>(Ljava/lang/String;)V at line 55
at [0x684fb914] Lorg/jikesrvm/adaptive/controller/ControllerThread; <init>(Ljava/lang/Object;)V at line 52
at [0x684fb914] Lorg/jikesrvm/adaptive/controller/Controller; createControllerThread()V at line 252
at [0x684fb950] Lorg/jikesrvm/adaptive/controller/Controller; boot()V at line 177
at [0x684fb990] Lorg/jikesrvm/VM; finishBooting()V at line 424
at [0x684fb9c0] Lorg/jikesrvm/VM; boot()V at line 157

This seems to have gone away in the last two regression runs, but I want to leave it open until I can have a chance to force GC to happen in exactly this place again by injecting an explicit System.gc() at the right program point. I'm suspicious that there is a GC map bug here that we just weren't hitting before, we hit it for a few sanity runs, then we stopped hitting it again.

By applying this patch to a development image, I get a 100% reproducible crash in GC during VM bootup

Index: /home/dgrove/rvm-trunk/rvm/src/org/jikesrvm/scheduler/RVMThread.java
===================================================================
— /home/dgrove/rvm-trunk/rvm/src/org/jikesrvm/scheduler/RVMThread.java (revision 14600)
+++ /home/dgrove/rvm-trunk/rvm/src/org/jikesrvm/scheduler/RVMThread.java (working copy)
@@ -410,6 +410,8 @@
contextRegisters = new Registers();
exceptionRegisters = new Registers();

+ if (VM.fullyBooted) System.gc();
+
if(VM.VerifyAssertions) VM._assert(stack != null);
// put self in list of threads known to scheduler and garbage collector
if (!VM.runningVM) { @@ -466,6 +468,7 @@ // create wrapper Thread if doesn't exist this.thread = java.lang.JikesRVMSupport.createThread(this, name); }
+ if (VM.fullyBooted) System.gc();
}
}

Have to stop for the day. I've traced it down to what looks like fairly convincing evidence that the call to AddressArray.create in JNIEnvironment.initializeState causes the TIB of the contextRegisters object of another thread (ie, not the executing one) to get smashed, even though the objects don't appear to overlap. This happens on a prototype-opt image.even if recompilation is disabled (as long as profiling is enabled, so we create the organizer thread).

./dist/prototype-opt_x86_64-linux/rvm -X:aos:enable_recompilation=false -X:aos:gather_profile_data=true -X:verbose -verbose:gc

With this instrumentation (plus some more higher in the call stack).

protected void initializeState(RVMThread target ) { if (VM.runningVM) VM.sysWriteln("initializeState 1 ", Magic.objectAsAddress(target.contextRegisters), " TIB ", Magic.objectAsAddress(ObjectModel.getTIB(target.contextRegisters))); if (VM.runningVM) VM.sysWriteln("this is ", Magic.objectAsAddress(this)); JNIRefs = AddressArray.create(JNIREFS_ARRAY_LENGTH + JNIREFS_FUDGE_LENGTH); VM.sysWriteln("Allocated JNIRefs array is ", Magic.objectAsAddress(JNIRefs)); VM.sysWriteln("sized ", JNIREFS_ARRAY_LENGTH + JNIREFS_FUDGE_LENGTH); VM.sysWriteln("check size ", JNIRefs.length()); if (VM.runningVM) VM.sysWriteln("initializeState 10 ", Magic.objectAsAddress(target.contextRegisters), " TIB ", Magic.objectAsAddress(ObjectModel.getTIB(target.contextRegisters))); JNIRefsTop = 0; JNIRefsSavedFP = 0; if (VM.runningVM) VM.sysWriteln("initializeState 20 ", Magic.objectAsAddress(target.contextRegisters), " TIB ", Magic.objectAsAddress(ObjectModel.getTIB(target.contextRegisters))); JNIRefsMax = (JNIREFS_ARRAY_LENGTH - 1) << LOG_BYTES_IN_ADDRESS; alwaysHasNativeFrame = false; if (VM.runningVM) VM.sysWriteln("initializeState 100 ", Magic.objectAsAddress(target.contextRegisters), " TIB ", Magic.objectAsAddress(ObjectModel.getTIB(target.contextRegisters))); }

I see the appended output.

allocateEnvironment 1 0x9f80014c TIB 0x600c3df4
allocateEnvironment 10 0x9f80014c TIB 0x600c3df4
the allocated environment is 0x70016318
initializeState 1 0x9f80014c TIB 0x600c3df4
this is 0x70016318
Allocated JNIRefs array is 0x9f800018
sized 150
check size 150
initializeState 10 0x9f80014c TIB 0x00000000
initializeState 20 0x9f80014c TIB 0x00000000
initializeState 100 0x9f80014c TIB 0x00000000
allocateEnvironment 100 0x9f80014c TIB 0x00000000
NEW THREAD cREG P30 0x9f80014c TIB 0x00000000

It's 100% deterministic at least, and seems to be not impacted by adding more and more sysWrites....

This error was introduced with the read barrier branch – sorry for the time it took tracking it down.

The register fields are explicitly traced by GC through the thread table, so we must keep the new register objects alive until the thread slot has been assigned.

Fixed in 14644.

reopening so I can modify fix target to 3.0