[9] RFR (S): 8173151: Code heap corruption due to incorrect inclusion test

Zoltán Majó zoltan.majo at oracle.com
Mon Feb 6 17:18:42 UTC 2017

Hi Tobias,

thank you for the feedback!

On 02/06/2017 04:52 PM, Tobias Hartmann wrote:
> Hi Zoltán,
> good catch!
> Here are some questions/suggestions:
> - Wouldn't it make sense to include an upper bound check in the guarantees as well?

That certainly makes sense, please see the updated webrev.

> - I wonder why AOT changed the code to use cb->code_begin() instead of cb, maybe someone more familiar with the AOT implementation could clarify this?

As we discussed offline, it seems that AOTCompiledMethods are allocated 
on the C heap (instead in the code cache). That is, only the code of an 
AOT method is allocated in the code cache. Could somebody from the AOT 
team please confirm that?

If the above is true, we could add a virtual method 
CodeBlob::blob_begin() to CodeBlob, which would return 'this'. 
AOTCompiledMethod would then override blob_begin() to return 
code_begin(). In CodeCache::get_code_heap() we would then use 
...contains(cb->blob_begin()). Would that make sense? (I'm sure we can 
find some better for that method.)

> - I'm not sure if we should make CodeCacheMinBlockLength a product flag because we would then also need more testing with different flag values. What about making it diagnostic or experimental?

Let's make it diagnostic then.

Here is the updated webrev:

JPRT testing is in progress.

Thank you!

Best regards,


> Thanks,
> Tobias
> On 06.02.2017 16:29, Zoltán Majó wrote:
>> Hi,
>> please review the fix for 8173151.
>> https://bugs.openjdk.java.net/browse/JDK-8173151
>> http://cr.openjdk.java.net/~zmajo/8173151/webrev.00/
>> The crash reported in the bug is caused by the corruption of the 'non-profiled nmethods' code heap. CodeHeap::_freelist for that code heap points to an address one segment before the heap's address space. The sweeper starts iterating through the code heap from the beginning of the heap's address space. Thus, the sweeper assumes that the first item in the code heap is a HeapBlock/FreeBlock (with the appropriate length and usage information). However, that is not the case, as the first item in the heap is actually *before* the heap. So the sweeper crashes.
>> This is a hard-to-reproduce problem (I managed to reproduce it only once in 350 iterations, each iteration taking ~25 minutes). So the fix I propose is based on core file debugging and source code investigation. But I managed to write a regression test that triggers a problem similar to the original problem.
>> I think that problem happens because a CodeBlob allocated in one code heap (A) is returned to a different code heap (B). When the CodeBlob is returned B, it is added to B's freelist. However, as the CodeBlob was allocated in A, the freelist of B now points into A (i.e., B is corrupted).
>> The code cache uses CodeCache::get_code_heap(const CodeBlob* cb) to determine to which code heap a 'cb' is supposed to be returned to. Since 8171008 (AOT) [1], the check is:
>> CodeHeap* CodeCache::get_code_heap(const CodeBlob* cb) {
>> assert(cb != NULL, "CodeBlob is null");
>> FOR_ALL_HEAPS(heap) {
>> - if ((*heap)->contains(cb)) {
>> + if ((*heap)->contains(cb->code_begin())) {
>> return *heap;
>> }
>> }
>> The blob 'cb' can be returned to the wrong heap if, for example:
>> - 'cb' is at the end code heap A and
>> - the size of the code contained in 'cb' is 0 (therefore code_begin() returns the address after 'cb', i.e., the first address in code heap B).
>> The fix proposes to restore CodeCache::get_code_heap() to its pre-AOT state (as I'm not aware of the reasons why AOT changed that check). I also propose to add some guarantees after allocation/deallocation in the code heap to possibly easier catch this or related problems in the future.
>> The regression test I propose achieves the above condition and results in a crash. The regression test works only with product builds, because in a product build a BufferBlob fits into one segment whereas in a fastdebug build it does not.
>> The test needs to set the CodeCacheMinBlockLength flag to 1. The flag is currently develop and we would need to make it product for the test to work. (Other flags controlling the code cache, e.g., CodeCacheExpansionSize, are also product.) I could also experiment with reproducing the problem with different block lengths/segment sizes, but that would most likely make the test more fragile (and CodeCacheSegmentSize is anyway develop as well).
>> I tested the fix with JPRT, RBT is in progress.
>> Thank you!
>> Best regards,
>> Zoltan
>> [1] http://hg.openjdk.java.net/jdk9/jdk9/hotspot/rev/777aaa19c4b1#l116.71

More information about the hotspot-compiler-dev mailing list