Native stack overlapping system heap base address on some systems

[Mike Hearn]

Hi,

Bug #2741 comes from a user who was able to grab a dump of the VM mappings
when the following error occurred:

  HEAP_CreateSystemHeap system heap base address 0x80000000 not available

I know people have been seeing this intermittently for a while now, we
have a few CodeWeavers support tickets with this problem too. The issue
seems to be that the Linux native (main thread) stack was allocated here:

   7fffb000     20K rwx--    [ stack ]

Surprise surprise, 0x7fffb000 + (1024 * 20) == 0x80000000

Unfortunately I lack enough knowledge about the kernels VM algorithms to
say why it's being allocated here. For comparison on my Fedora Core 3
machine:

   fef0d000     16K rwx--    [ stack ]

One other thing I don't understand. It seems the stack and the system heap
overlap by exactly 1 byte, otherwise the mmap would not fail. But, this
area is (theoretically) being protected by the preloader: so how can this
be? There is something I'm missing here, I suspect an off-by-one error in
my arithmetic.

thanks -mike

[Alexandre Julliard]

Mike Hearn <mh(a)codeweavers.com> writes:
> I know people have been seeing this intermittently for a while now, we
> have a few CodeWeavers support tickets with this problem too. The issue
> seems to be that the Linux native (main thread) stack was allocated here:
>
> 7fffb000     20K rwx--    [ stack ]
>
> Surprise surprise, 0x7fffb000 + (1024 * 20) == 0x80000000
>
> Unfortunately I lack enough knowledge about the kernels VM algorithms to
> say why it's being allocated here. For comparison on my Fedora Core 3
> machine:
>
> fef0d000     16K rwx--    [ stack ]
>
> One other thing I don't understand. It seems the stack and the system heap
> overlap by exactly 1 byte, otherwise the mmap would not fail.

No it doesn't overlap, a 20K mapping at 7fffb000 ends at
7fffffff. Most likely that kernel is configured for a 2Gb virtual
address space.

-- 
Alexandre Julliard
julliard(a)winehq.org