svn commit: r280327 - in head/sys: kern vm

[Don Lewis]

On 28 Mar, Alan Cox wrote:
> On 03/28/2015 12:29, Bryan Drewery wrote:
>> On 3/27/2015 9:41 PM, Don Lewis wrote:

>>> I was not seeing this problem on my older package builder running
>>> 10.1-STABLE.  Since this problem has not shown up on the FreeBSD package
>>> building cluster, I got suspicious that the change was quite recent.
>>>
>>> This old gcc bug report:
>>> <https://gcc.gnu.org/bugzilla/show_bug.cgi?id=14940> led me to suspect
>>> mmap().
>>>
>>> The old gcc source file /usr/src/contrib/gcc/ggc-common.c does a couple
>>> of mmap() calls.   Tne first is in mmap_gt_pch_get_address() where a
>>> NULL first argument is used.  The address that gets returned is stashed
>>> away and the region is unmapped.  Then a later call in
>>> mmap_gt_pch_use_address() passes this saved  address to mmap() as a
>>> hint.  It expects the mapped region to get mapped to the same base
>>> address.  If this does not happen, the above error is the result.
>> I don't know what I'm talking about but that doesn't sound like a very
>> safe assumption for the gcc code to make.
> 
> 
> Your intuition is correct.  It is not.  And, in fact, under Solaris and
> Linux, gcc does not make this assumption.
> 
> I suspect that the solution used by gcc under Solaris would work for
> us.  However, that would entail modifying gcc in older branches.

That's somewhat problematic because we use releases + official updates
for package building.  To continue doing this, we'd have to treat the
gcc fix as an EN.