Review request (S): 7021322: assert(object_end <= top()) failed: Object crosses promotion LAB boundary
y.s.ramakrishna at oracle.com
Mon Sep 12 17:14:30 PDT 2011
Good catch. Fix looks good.
I am also a bit leery of any unallocate failing. It seems to me that
should provably succeed. Thus the "return false" path seems bogus to me,
(in my book) be replaced with either an assert(false) or ShouldNotReachHere.
Then the object filling code becomes redundant and can be removed.
On the other hand, may be I am missing some reason why the unallocate might
legitimately fail and the object filling becomes necessary in that case.
Looks good otherwise.
On 9/12/2011 1:49 PM, Stefan Karlsson wrote:
> There's a bug in parallel scavenge when array chunking is used. After
> a thread has succeeded in forwarding the pointer to a newly copied
> array, it might change the length of the old object. This is done as a
> part of the load balancing.
> If another thread races with the forwarding thread, it might read the
> incorrect array length and copy just a part of the array. When it
> later sees that the object has already been forwarded and calls
> unallocate_object, it uses the original length of the array to
> determine the size to unallocate.
> The fix is to pass the actual amount of memory that was allocated, to
> Tested with the failing test.
More information about the hotspot-gc-dev