[Mono-dev] Issues with GC due to libgc

[Dick Porter]

On Wed, 2009-09-30 at 10:24 +0200, David Suarez wrote:
> Hi Zoltan, 
> 
> Let me point some data from our tests to the discussion. 

> Those blocks up to 100Mb are actually "leaked" I'm afraid. If you do
> again the same operation in the sample (allocate 2.5GB), your virt
> memory will again go up to 2.5GB + the "free" blocks, so they seem to
> be not so free.

To follow up from David, the amount of 'leakage' is also variable from
run to run.  If it was just plain fragmentation I would expect the
memory usage to be at least consistent.

We are also seeing that some objects are being allocated without a GC
descriptor.  I noted Zoltan's response about the GC_GCJ_MALLOC functions
and need to read the code again.

We also have a theory that with libgc the false pointers on the stack
aren't being removed from the mark stack when the execution stack
unwinds.  I will be investigating this.

- Dick