[Omake] SVN Commit: OMake Build System [0.9.8.x] (Rev. 12456)

[Aleksey Nogin]

On 30.10.2007 12:47, Jason J. Hickey wrote:

> I believe I have solved the problem with bogus termination on Windows XP.
> See the message titled "Parallel builds on windows xp" from Doug Rabson on
> Aug 9, 2007.
> 
> I *can't* reproduce the problem, but I have an explanation.
> 
> If correct, the problem was a double-allocation in omake_shell_sys.c:handle_wait,
> where
>    1.   tuple = alloc_tuple(2);
>    2.   Field(tuple, 0) = ...
>    3.   Field(tuple, 1) = Val_unit;
>    4.   status = alloc_small(...)
>    5.   Field(tuple, 1) = status;
> 
> Line 5 violates the assignment policy, which says that direct
> assignment is allowed only if the tuple was just allocated.
> 
> So, the explanation would be that a GC happens at line 4, moving the
> tuple to the major heap, and the assignment on line 5 violates the
> generational invariant (by not keeping track of the old->young pointer).
> This leads to heap corruption, and the result looks like "process stopped",
> only because the stopped case is the last one in pattern matching,
> so it serves as the default.

Makes sense. And note that for the corruption to happen there needs to
be another GC before the value is consumed (which happens quite quickly)
- so this will probably require another memory-hungry thread intervening
in parallel. This would explain why Doug is the only one seeing this -
there must be something fairly unique in his setup, since hitting this
bug requires quite an unusual set of circumstances.

Ah, I wish there was a better way to make sure the C code is not
breaking things. I am sure it took you quite a while to find this :-(

We should probably ask Doug to try out the latest code to see if this
have indeed solved his problem.

Aleksey

-- 
Aleksey Nogin, Research Scientist
Advanced Technologies Department, Information & System Sciences Lab
HRL Laboratories, LLC, Malibu, CA