Patch: PR 16748 (Re: hash synchronization on linux)

[Bryce McKinlay]

The attached patch should correct this problem, using Hans' first 
suggestion below. I'm going to check this in to HEAD and the GCC 3.4 
branch, and maybe GCC 3.3 branch as well.

Thanks to you both for tracking this down.

Regards

Bryce


Boehm, Hans wrote:

>The problem is in thread 8.  I doubt it has been fixed,
>but I didn't check.
>
>Basically we have
>
>_Jv_MonitorEnter
>	acquires lock bit on lightweight lock entry
>	notices it needs to allocate heavyweight lock
>	calls eventually alloc_heavy.
>alloc_heavy call GC_local_gcj_malloc
>GC_local_gcj_malloc is out of free list entries AND
>	is first allocation to run since GC, which
>	results in a call to
>FinalizerThread.finalizerReady, which is currently
>	unrestricted Java code.  It then calls
>_Jv_MonitorEnter
>	which tries to acquire the same lock table
>	entry it already owns ==>
>
>deadlock
>
>Is it hard to make FinalizerThread.finalizerReady a small piece
>of C++ code, which does not
>acquire Java locks, e.g. by just doing a pthread_cond_signal
>equivalent?
>  
>
>The alternative would be to avoid allocation calls with the
>lock bit on a lock table entry set.  (I think that registering
>a finalizer is OK.  That allocates internally, but I don't
>think it can invoke finalizer notification.)  But that looks
>a bit messy and more expensive:  You'd have to preallocate the
>heavyweight object before acquiring the lock bit, and then drop
>it if you didn't need it.
>
>Or I guess we could add a GC allocation call which didn't do the
>finalizer notification check.
>
>I'm also not sure that running Java code during finalizer
>notification is otherwise safe.  This can happen during arbitrary
>allocations, with locks held, etc.
>
>Hans
>  
>
>>-----Original Message-----
>>From: java-owner@gcc.gnu.org 
>>[mailto:java-owner@gcc.gnu.org]On Behalf Of
>>Jacob Gladish
>>Sent: Thursday, July 08, 2004 2:32 PM
>>To: java@gcc.gnu.org
>>Subject: hash synchronization on linux
>>
>>
>>I've suspected a problem with the hash synchronization for 
>>some time now
>>and decided that I might be able to reproduce the problem with a unit
>>test, and think that I have stumbled onto something. 
>>I've attached two classes that can be run against gcc-3.1 and very
>>quickly deadlock. To make this happen you must rebuild with a changed
>>hash_entry table size of 2. I suspected a race with bucket 
>>access (heavy
>>lock chain) so reducing the size increased the likelihood that I would
>>be able to reproduce this easily. This is being run on Linux 
>>SMP 2.4.18
>>(machine has two procs).
>>I've also included the backtrace of all the threads in this state. No
>>threads ever return from _Jv_MonitorEnter. I know the 3.1 release is
>>pretty old at this point, but I'm in the middle of building 3.4.1 with
>>the hash_entry table size shrunk to 2 and am going to see if 
>>it behaves
>>the same. I was also under the impression that the 
>>hash_synchronization
>>code has basically been untouched for a while.
>>Thanks
>>Jake
>>
>>    
>>

-------------- next part --------------
A non-text attachment was scrubbed...
Name: gcj-PR16748.patch
Type: text/x-patch
Size: 5340 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/java/attachments/20040711/dcaaa878/attachment.bin>